Vibecoding general - How to become the a 10x engineer you always knew you were despite being absolutely inept in every imaginable manner

  • 🏰 The Fediverse is up. If you know, you know.
  • Want to keep track of this thread?
    Accounts can bookmark posts, watch threads for updates, and jump back to where you stopped reading.
    Create account
have you tried vibe avellooning? image(2).png
 
Is there any decent tooling for doing this locally without connecting to the Internet?
I know local LLMs aren't as good in general, but I am extremely wary of becoming dependent on something that can be taken away from me.
 
Squishie PP said:
Is there any decent tooling for doing this locally without connecting to the Internet?
I know local LLMs aren't as good in general, but I am extremely wary of becoming dependent on something that can be taken away from me.
Click to expand...
You need about 128GB+ of RAM minimum to have a good local coding agent.
 
debristar said:
there's no such thing as vibe coding. karpathy is too much of a coward to say he made a mistake letting it catch on. that in reality it should have been called hybrid programming.
Click to expand...
because at first vibecoding literally was vibing with coding
you wanted to make yourself a little homepage or something that didnt require that much thought? sit back and vibe and watch ai code
only after that it started being used by INCOMPETENT RETARDS (and jeets) to spit in the face of computer science more than webdevs did by having 100000GB node_modules
PikminCollector said:
It depends on how you direct it I suppose - There's lots of market hype over Mythos right now, which is allegedly highlighting Day 0s faster than anyone can keep up.

Click to expand...
(im writing this reply for the third time now, for the first time i accidentally refreshed the page then i clicked cancel instead of save)
mythos isnt really all that great (kind of like how every single big llm now is a phd level mathematician and shit)
in curl for example mythos claimed to have found "confirmed five security vulnerabities, when humans have taken a look at it it turned out that only one of them was a low severity vulnerabity, three others were false positives and another one was "just a bug"
 
Last edited:
Andy Soying Out said:
You need about 128GB+ of RAM minimum to have a good local coding agent.
Click to expand...

I've heard qwen3.6:35b-a3b and kimi-2.5 are both pretty decent for local models, but I haven't tried either. qwen won't load on my 32GB AMD workstation card via ollama. Looks like it needs somewhere around 50GB~64GB?

Squishie PP said:
Assume that isn't a problem. Also I am not asking about which specific model weights to use, but inference software with good integration into IDEs.
Click to expand...

I've used Continue before with Jetbrains IDEs (IntelliJ, Pycharm, etc.) and it works decently well. It has a similar UI to the copilot or claude plugins and can connect to an ollama server where you can serve models from.
 
Breadbassket said:
How vulnerable is Vibecoded programming compared to manual human programming I wonder?
Click to expand...
Keep in mind Ai is trained on mostly junior code and free example/portfolio projects, so while it might be quick and meet your specifications, it has the combined security knowledge of all the tech juniors and students on Github.
 
Overall vibecoding works pretty well as long as you are just using it to handle tedious tasks and don't have high expectations.

The Enola Gay said:
Nigga I have a codebase that consists of over 400 php files. I'm not copypasting code files one-by-one out of my browser into the terminal.

Can confirm, Opus is garbage. Not worth the extra token consumption at all.
Click to expand...
I'll second that. I had Claude Sonnet take a screenshot and a proprietary XML markup file and render a one page HTML5 app out of it.
Opus blew 10 times as many tokens to spit out an HTML file that didn't fucking work. Pathetic.

I got a few months of Gemini Pro free through work, and the Google AI Studio makes some pretty competent stuff... as long as it's TypeScript, React, and uses Firebase... and you don't mind running your app in Google Cloud. I suspect AI Studio is just a way for Google to push people to their other products.
 
Mad as a Swan said:
Keep in mind Ai is trained on mostly junior code and free example/portfolio projects, so while it might be quick and meet your specifications, it has the combined security knowledge of all the tech juniors and students on Github.
Click to expand...
That isn't really true. I had OpenCode (just using their Big Pickle model) and Claude essentially fix a bunch of vulnerabilities and put in scanning (via CI) into my project. It will also inspect library sources and build relatively ok code most of the time.
 
Vibecoding is great if you tell it more than just “make gaem pls” and actually know how your program should be structured but just don’t want to mess with getting the syntax right. It’s both impressive and horrifying knowing I could now take a few hours to shit out programs that once took me a week to make, and the AI version would be better.
 
Two Dollars said:
Vibecoding is great if you tell it more than just “make gaem pls” and actually know how your program should be structured but just don’t want to mess with getting the syntax right. It’s both impressive and horrifying knowing I could now take a few hours to shit out programs that once took me a week to make, and the AI version would be better.
Click to expand...
All automated systems are inherently complex, and complexity increases the chance of failure. LLMs are no different. It is still important to be vigilant about understanding the code they spit out.
 
Squishie PP said:
All automated systems are inherently complex, and complexity increases the chance of failure. LLMs are no different. It is still important to be vigilant about understanding the code they spit out.
Click to expand...
True, you don’t need to know how to write a lot of functions anymore, but you still need to know enough to look at a generated function and know if it’s right. You don’t need to remember the syntax of how a pointer is used, but you still need to know to tell it to use pointers.
 
Two Dollars said:
and the AI version would be better.
Click to expand...
Are you people really not looking at the code? Even with 4.7 Opus and Sonnet models, with well written plans and rules, I still see these models outputting horribly convoluted series of nested try/loop/conditional blocks, with prompts for refactoring usually making things worse.

The models also like to do blanket catch-all exception blocks, in every language. That swallows up errors silently, making the output seem like it work more correctly than it actually does.

It is impressive so much of it works, but when it gets it wrong, it's really wrong.

I feel like no one cares about maintainability anymore. We had a lead engineer leave at work and I assigned to be the lead on the project. He loved Claude, and it shows. Today I found some code that ignored an existing GUID and always generated a new one when writing a record. I find stuff that never should have made it past code review or QA every day.
 
HootersMcBoobies said:
Are you people really not looking at the code? Even with 4.7 Opus and Sonnet models, with well written plans and rules, I still see these models outputting horribly convoluted series of nested try/loop/conditional blocks, with prompts for refactoring usually making things worse.

The models also like to do blanket catch-all exception blocks, in every language. That swallows up errors silently, making the output seem like it work more correctly than it actually does.

It is impressive so much of it works, but when it gets it wrong, it's really wrong.

I feel like no one cares about maintainability anymore. We had a lead engineer leave at work and I assigned to be the lead on the project. He loved Claude, and it shows. Today I found some code that ignored an existing GUID and always generated a new one when writing a record. I find stuff that never should have made it past code review or QA every day.
Click to expand...
If you want it to make great code that matches your exact preferences, you should tell it what coding practices to follow, point it to what already exists in the codebase so it knows what to use, write tests and use them to verify whatever it makes, like a person. Or as I already said:
Two Dollars said:
Vibecoding is great if you tell it more than just “make gaem pls” and actually know how your program should be structured but just don’t want to mess with getting the syntax right.
Click to expand...
 
You must log in or register to reply here.
Back
Top Bottom