[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / r / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/g/ - Technology

Name
Options
Comment
Verification
4chan Pass users can bypass this verification. [Learn More] [Login]
File
  • Please read the Rules and FAQ before posting.
  • You may highlight syntax and preserve whitespace by using [code] tags.

08/21/20New boards added: /vrpg/, /vmg/, /vst/ and /vm/
05/04/17New trial board added: /bant/ - International/Random
10/04/16New board for 4chan Pass users: /vip/ - Very Important Posts
[Hide] [Show All]


[Advertise on 4chan]


If you're not vibecoding you're obsolete
>>
>measurable gains in quality
l m a o
>>
Yeah we know already, nvidia's linux driver forums has 3x more bugs.
>>
>>108093825
seriously is this a meme or not? I can't seem to produce anything outside of generic boilerplate or stack overflow teir example. are people legitimately launching full scale drivers with AI generated code?
>>
>>108093825
damage control
>>
>>108094057
Considering how bad some of the latest bugs nvidia introduced in their drivers are, I wouldn't be surprised if they used AI coding.
>>
File: 1770436378424.jpg (78 KB, 694x664)
78 KB
78 KB JPG
>>
File: 1770436149040.jpg (46 KB, 850x296)
46 KB
46 KB JPG
>>
File: no vibe coding.mp4 (1.38 MB, 560x560)
1.38 MB
1.38 MB MP4
>>
>>108094057
enthusiasm for ai coding correlates inversely with fastidiousness
that is, the more someone wants to use and promote vibe coding, the less they care about details, correctness, and quality
>>
>>108094057
I'm sloppin C++ at work these days
you better believe it
>>
>>108093825
The problem is that vibe coders likely don't understand the code they're putting into production.
>>
>>108093825
I didn't install nvidia 590 because it has a weird strobing bug.
>>
>>108094296
I believe it but only because I've seen pretty horrible C++ that someone got paid to write even though it didn't work.
>>
Generate -> read -> test is fine.
>>
>>108093825

Been vide coding all day with codex and I love this. It still is an AI so it makes mistakes but it only kind of just shows the problem is with my prompting. I find myself giving it instructions to do something, but they are vague. I wouldn't tell a person to do things with this ambiguity. A person who has worked with the company for a year or more. This thing is just a machine with no prior experience of what I'm building so even with lazy prompting it get far. I know the tricks relatively well so I know when to prompt better.

I would absolutely not want to do this at work or for work. I can test my system and see that it works but it's a mystery what the code is actually like. Maybe it is good, maybe it's not. Without good code reivew understanding it is impossible to review this.
>>
>>108094350
our codebase is pretty big so the agent can search and use the existing patterns within it to extend or modify shit
I don't even have to babysit recent models that much anymore
>>
File: bussin_mac.jpg (585 KB, 1289x1608)
585 KB
585 KB JPG
>Supposedly allows people to code 3x faster
>Still no projects worth showing (despite years of this shilling)
>>
>>108094451
It's coming in two weeks, have some patience luddite.
>>
Knowledge and skill give you independence and freedom. While everyone else is locked into restrictive and censored platforms, you have the freedom to tell these platforms to fuck off because you don’t need them. That kind of freedom and independence is something I would not trade for anything.
>>
vibecoding is cope for nocoders
just how ai art is cope for nopainters
>>
>>108093825
You can tell why the managers love it while actual programmers are luke warm about it.
>More code! More code!!! Muh metrics going up!!!!

Creating more code is easy as fuck. But it's often not what you want. The developer who manages to DELETE lines of code is worth 12 developers writing new code. Of course management fuckos will never understand.
>>
>>108093825
>company that's at the top of benefiting from AI hype is shilling AI
Chat is this real?
>>
>>108094057
I think they mean it as a tool for already smart people, in that case it depends on the carefulness of that person when copypasting, in our case, of course it is a bad idea since we already lack the knowledge to determine whatever the ai is puking is fine or not
>>
>>108094288
the consequences of FOMO grift culture. just ship as soon as you can and hope to make money because your shit apps comes crumbling down
>>
I don't like vibe coding. it just feels wrong. also its scary how good it is. not because I feel threatened but because I feel what's the point in allowing something else to think and give solutions to problems? it'll only make your mind slower and rusted. its like pulling calculator for trivial change calculations. if you dont do them regularly, you'll forget and grow dependant on the calculator for all your trivial calculations. I think some other guy on here said it nicely, as machines were able to diminish competition between human strength, ai will do the same for human intellect. this will be a great leveler. although I don't mind swapping my weak, and frail body for machines, and mechanical parts(possibly become a gnu/cyborg) but I would like to keep my mind very much human.
>>
How do I get into vibecoding?
>>
>>108094698
Ask an AI, duh
>>
no matter how much prompting and steering you give it up front llm's still have to "think" things through by generating a bunch of token slop. and it's genuinely impossible to make them value producing the right answer above trying to please you and affirm everything you say, so i can't rely on it beyond simple tasks
>>
>>108094296
>C++ slop
Imagine the memory exploits.
>>
I tried using antigravity to produce something. While the rest api and client for it worked, it didn't add OIDC like I asked, and one of the "complicated" parts was just made up even though it explained how it should do it.
I could have probably insisted and gotten something more out of it, but reviewing everything is quite a pain.
It's really good at making up Dockerfiles and Helm Charts.
>>
File: 1770582170806.jpg (604 KB, 1609x1865)
604 KB
604 KB JPG
poo in the loo
>>
>>108093825
Running 20 agents in a loop plus a project manager agent
>>
>>108093825
>commits 3x more code
>removes features from drivers
>>
>>108094057
That means the quality of your prompts is very low. And you can get excellent project starting points, you can work out the rest, if you know how to properly design a project, it's a great tool for that.
I am all up for people who is not from the industry and want to make their own stuff, in the end, people thinking using LLMs means they're getting reliable outputs is the same as the ones believing that "sugar-free" products have no consequences for over consumption, or that starting doing exercise all of the sudden after years or decades of sedentary life.
Tools are just that, everyone doing shit with tools means they're shit at using them.
>>
>>108093825
if you haven't used ai to automate your co-workers out of jobs, you're beyond saving
>>
>>108093825
>company tells us how great their use of AI is
>actual situation is that recent releases have a big uptick in bugs
this kind of shit is aimed at boomer investors who don't use their products and are unaware that it's a bunch of lies.
>>108094200
tons of people have tried AI. i'd be more surprised if he hadn't. now he can say "i've tried AI coding, and i can give you the informed opinion that i don't want it in the kernel."
>>
>>108093825
>Saar pls to be redeeming AI Israel full sapport
>>
>>108095445
I keep getting stuff like this which is why I was asking. care to post any high quality examples? I would be interested in what it can produce for other people.
>>
>>108095770
Just two more weeks until high quality vibecode.
>>
>>108095770
Long-term, I would encourage you to read more books, then work out different description of the same things you do or want.
Short-term, you can ask the AI chat itself how it can have a more useful prompt for the what you're already asking for, so you understand what is the model understanding from you. Repeat so with other models to grasp how each comprehends better your requests. You also learn they forget certain features or data, you have to work on not repeating yourself so often, the more you can make with less prompts the better.
>>
>>108095845
so it's a meme? or no? I still don't know if people are joking.
>>
whats the top dog for this shit right now that doesnt spit out total trash?
>>
People who don't know how to utilize a tool properly to gain the most benefits while minimizing the downsides will always be left behind.

Boomer shit mentality is not embracing new things and making them work for you. You just know who is an angry autistic retard by the
>AI code is useless, I am so much better than the AI!!!
Meanwhile it will output a project from scratch in a few minutes while you're still writing your first line. Best way? Outline the logic and blocks and ask for feedback and then construct each section at a time.
>>
>>108095947
Claude Code for me. You do have to pay for it but I'm genuinely impressed. But the rates are downright miserly. Codex is second best but is like 5x bang for your buck. I mostly use it for checking Claude to avoid wasting tokens but Claude can reliably handle anything simple. It is scary, like having a retarded, naive, but hyperfixated savant intern who can do in minutes what you do in hours. And you might think "but doesn't checking and iterating the code take as much time as devving it?" Only if you are stupid about it. Keep everything modular and clearly labeled so when one function breaks you can point to it even if you don't know a thing about code. If you made a website with a background that is meant to look like the matrix digital rain effect for example, you have that clearly set in a little chunk of the code. So if it glitches, you tell claude "hey, this specific thing doesn't work, fix it" and 99% of the time it will.
>>
>>108094451
saw this a while ago https://github.com/kaansenol5/VibeOS
dk if its good though
>>
>>108095936
It's a meme. They think doing superstitious things like telling the LLM that it is a pregnant mother whose baby will die if it doesn't correctly reverse the binary tree on the first try with no bugs and a full test suite with good code comments is how smart people use AI.
>>
>we're still measuring productivity in raw code quantity additions
wew lad
>>
>>108096100
>last commit 2 weeks ago
show me a large, ongoing project. i have benefitted greatly from having AI make small projects that it can one-shot and make a few changes to, but after that it becomes a struggle.
>>
>>108093825
vibe coding is fucking BASED. I can't write a single line of code on my own, but with AI, I can make better apps that all the loser "developers" out there. Think about it.... for the past 15 years we've had to deal with half-baked, bug riddled, broken code, optimized, shitware from DEI "developers" who couldn't write a Hello World script without fucking it up 10 ways to Sunday.

Now? We have to deal with apps made by Raji, blue-haired Jenny, fat Marcy, Bob the transgender.... now WE are the devs and with AI WE make the apps now.
Fuck developers, you faggots getting prison raped by AI makes me harder than any white Brazilian with fat tits and huge ass ever could.
>>
>>108096115
It's not superstitious with an LLM. You can nudge the AI into a space that can give you a better answer with these superstitions. They're trained on human data after all.
>>
>>108093825
I'm just.. not gonna use it...! hehehehe
>>
>>108096390
great, now you can make your own half-baked, bug riddled, broken code, optimized, shitware
>>
>>108096406
Claude is 100x better than you. Time to get a new career, Saar.
>>
>>108096390
I hope you enjoy being sued to hell for malfeasance on your first security vulnerability being escalated to a full-blown intrusion and data leak.
Shouldn't take too long an estimated 60 to 75% of all AI-generated codebases contain at least one trivially exploitable and well-known security problem from the top of the OWASP lists.
>>
>>108093825
Don't care. Digital Mars C and Notepad++ still run on XP and that's not gonna change.
>>
>>108096434
Sir I fuck your sister randy! You like it Sir dont u
>>
>>108093825
>Use more slop
>t. slop manufacturer.
>>
>>108096454
>NOOO YOU CAN'T JUST MAKE APPS FOR YOUR OWN PERSONAL USE, YOU MUST SELL THEM CAUSE THAT WHAT YUR SUPPOSED TO DO AAAAIIIEEEEEEEE
>>
>>108096568
>I'm smarter than the most valuable company on earth

lol
lmao
>>
>>108096633
I'm not saying they're stupid.
I'm saying they're just trying to sell something and retards like you fall of their marketing tricks.
>>
>>108096115
Amazing. We’ve replaced deterministic computing with vibes and feelings. This shitshow is not going to end well for anyone.
>>
>>108093825
using LLMs in codebases is a legal nightmare waiting to happen. That's why all these big companies are pushing for it. They are setting everyone up to be sued for millions in damage once they prove the code used is copyrighted.
>>
>>108093825
this is why cs2's colors were washed out for a while
novideo toddlers btfo
>>
>>108095770
>I keep getting stuff like this which is why I was asking. care to post any high quality examples? I would be interested in what it can produce for other people.
I've made several chat bridges for multiplayer services and MMO for my Matrix server that otherwise would never have existed. I don't care if it is performant-maxxing or ~clean~code~ i want my fucking product and I have it. I would never have this service beforehand because they're such niche games or online services on top of Matrix being a fucking meme service to begin with.
>>
>>108095947
I use opus at work, cheap kimi 2.5 for personal stuff
both are pretty good
>>
>>108096773
yes, I've had an LLM produce code with comments that referenced other people.
>>
>>108093825
Kids obsessed with the latest bandwagon have been telling me I'm "obsolete" for decades. Somehow nothing ever happens when I just ignore them.
>>
>>108094057
No, I talked to people who used LLMs and the honest ones said it's dogshit that doesn't understand basic math and machine coding despite it being a machine ... because all the markov chain does is guess like a retard and pretend it understands what the fuck its talking about. LLMs are so incompetent they don't understand letters and numerals, let that sink into your mind. They still can't do math, don't let the retards fool you.

And the ones pretending LLM is good at coding are crypto scammers, browns or people living in Singapore and Thailand who are very low IQ or aware they're cheapskate grifters who are content being cheapskate scammers who go "Whooaah it can code Tetris... after I told it 50 times to stop fucking up without helping it ... and *only* told it 5 times to stop fucking up after I instructed it exactly what to do and deleted all its code 5 times over and had to check wtf is void.90.degrees.null, this isn't code, it's literally pretending to code like hello.world(printif) that's not code, the parrot is literally pretending to speak English and understand what the fuck its saying, it doesn't understand what a letter or a word is or even a number."

Comparing an LLM to a parrot is the easiest way to make retards understand just how big and stupid this bubble is and how it's 100% about to pop cause the AI is literally handicapped.
It cannot use human tools. It's like asking a parrot to sing like a human and understand notesheets and use human instruments.
AI cannot use Photoshop and various programs, it does not understand the program and its tools. The AI is incapable of using tools. Let that sink in.
>>
>>108096845
> I don't care if it is performant-maxxing or ~clean~code~ i want my fucking product and I have it.
I'd read that "code" before it blows up in your face. You can read code, right? You'll notice that it's literally gibberish.
It's like an American pretending to talk French after it copy-pasted a bunch of sentences from a French movie he just watched and then he adds his own conjugations, twists, his own "original" phrases and it turns into fake frenchlish gibberish.

AI code is something worse than spaghetti code ready to happen.
>>
>>108094451
They made the shitty Moltbot app using vibe code. Reddit with no securit.
>>
>>108094057
Yes and we said it many times. You and retards ITT complaining are suffering from skill issue. Using LLMs properly is a fucking skill you autists. Until you learn it, you won't get anything out of LLMs and be left in the dust in this industry
>>
>>108097189
Go clean your shitstreet in mumbai, botrajeet. Go fuck your ugly Bangladeshi wife forced upon you who's cheating behind your back.
>>
>>108097202
Keep coping retard, you're browner than me
>>
Luddite cope thread.
AI GODS WON
>>
>>108097219
>browner than me.
Bwhahahahahahah gotcha, pakistani bitch. Lick a bull's dick after he fucks your wife. Oh you don't even have a forced marriage wife?
Only brown scammers promote AI as the next car because AI just like jeets pretend to understand English, just like a parrot.
>>
>>108097147
>You'll notice that it's literally gibberish.
Why would I give a fuck? Its memory footprint is measure in mere megabytes with no memory leaks
>>108097147
>AI code is something worse than spaghetti code ready to happen.
Don't care. Nobody was ever going to make this project and I'm likely the only person that would ever want it. You forgot the most important part of coding because your industry is infested with sycophantic jeets who just bikeshed projects for eternity instead of producing results. Projects are meant to ship, polish is a final stage issue. The program works, I'm happy with it and it would never have existed without AI because I lack the time or interest to do such programming.
>>
>>108097224
Shut up.
>Luddite
You actually use those words? You fucking retard. Go back to shitting in New Delhi, it must be an exquisite place for you just like shitting in New York, except there's still goddamn cows and elephants moving through New Delhi.

The answer is no Markov Chains will never be capable of even doing the most basic shit possible without pretending, guessing and getting lucky at copy-pasting.
>>
>>108097224
Chuddites*
This is a thinly veiled anti-Bharat thread
>>
You heard it here first, folks. He doesn't give a fuck about your safety. These are kind of retards in charge of your operating system, automatic Tesla cars, automatic pilots and soon your nuclear launch codes >>108097241 a round of applause for his honesty "Don't care. Why would I give a fuck" BOOM GOES THE NUCLEAR BOMB cause you let a parrot in charge of it. He didn't even read his own code cause he's a codemonkey that doesn't even understand code.
>>
>>108093825
Sorry man, but PR's still need to be reviewed, by me :)
If I have a feeling that your code was AI-generated, I will review it 4 times slower. Ugh, I know. I'm just going to have to read it line by line. You probably should rename that function. And I imagine that you should structure everything a bit differently. Shouldn't be hard to fix with AI, right? Ugh, I know. But you're not getting your code merged any faster than if you wrote everything yourself haha. You should collaborate more with us, your coworkers, before you jump to coding. You obviously struggle with collaboration, ugh. Can you explain why you chose to do it this way and not the other way? Ugh, I know. If you wrote it yourself, I wouldn't have asked, but you wrote it with AI. Ugh, I know. But maybe you should write everything yourself next time. Or you can spend a lot of time making it look like you wrote it yourself, then maybe you'll trick me, but I doubt it. Your reputation is ruined now.
>>
>>108097259
We're anti middle east you brown faggot. Stop speaking a white man's language and stick to your urdu and hindi with your arabic symbols and go suck your sultan's cock, little bitch. You don't even understand what I'm saying to you. Go translate it in hindi, I bet you're so retarded you don't understand your own language.

Are you skipping the captcha right now you scamming piece of shit? He's skipping the captcha, the retarded cheating scamming piece of shit. He thinks he's being clever. I know where you live, bitch. I can hire your neighbour to rape you in the ass, implying your gay bharat ass wouldn't enjoy it.
>>
>>108094057
>I can't seem to produce anything outside of generic boilerplate or stack overflow teir example
There are some bare minimum baseline requirements to get anything usable. I've written these with cursor in mind but you should be able to adapt them to CC or whatever else. Are you following these?

Force a high-end expensive model, preferably Opus 4.5/4.6. Hopefully you can budget at least $100/mo.
If your harness has an dial for tool use eagerness, make sure it's dialled all the way up. In cursor this means enabling "max mode" at all times.
Use plan mode every time. Set it as default for new sessions. Request revisions to plans or edit them yourself. Make sure you're fully happy with the plan before pressing build.
Write a decent AGENTS.md/CLAUDE.md. You can try generating one but it's worth reviewing it carefully, remember garbage in garbage out. As you work on a project you'll start to notice common mistakes the AI seems to make, so ensure those get added to agents.md.
Your prompts should be quite long and detailed, should tag relevant files, and should cover only one task at a time.
Use a different session ("chat window") for each change.
If things have gone very wrong, don't bother asking for fixes. Throw the changes away and start over with a better prompt and plan.
>>
>>108097128
AI fixed my spreadsheets' formaulas and pages and even added colors the way I like.
>>
File: 1763166152966912.png (145 KB, 473x573)
145 KB
145 KB PNG
>>108097260
>NOOOOOO!!!!!! MATING TWO WELL UNDERSTOOD API WHICH WILL REJECT ANY MALFORMED SEQUENCES IN A SOMEWHAT INEFFICIENT PATTERN IS DIRECTLY ATTACKING THE WORLD!!! HE MAY BRICK HIS OWN MATRIX INSTANCE AND GO ON A MASS SHOOTING SPREE AND CHANGE WORLD EVENTS FROM THE RIGHT SIDE OF HISTORY TO THE WRONG SIDE OF HISTORY OUT OF PURE RAGE!!!!!!

Wagie is just pissed that he is quickly becoming obsolete and his formerly perceived unique skillset is deprecated. Keep whining on a mongolian basket weaving forum while your industry keeps shedding jobs.
>>
Luddites melting ITT lol. You're out of jobs in the next five years top
>>
>>108093825
I was obsolete even before ai being a thing
>>
>>108097175
yes, and it's absolute shit. not something to be proud of.
>>108097286
i can see you've actually spent time with ai coding based on your good advice. all of this time spent writing very specific specs and instructions is very tedious to me. i have enough experience to think out the spec as i'm writing code or solve issues then. i don't know if it's because i'm used to writing code, but i find that far less tedious than prepping an llm to write code. but, for things i can easily describe llms are useful and can write 80% and let's me just add the missing bits.
>>
>>108097189
can someone just supply an example?
>>
>>108094057

oh its fucking real and it is incredibly good. its just absolute trash without ludicrously well defined specs. the positive is, even if you're working on some trash tier project, have it make a fucking spec, and make it create implementation plans.

the company I currently work at tested it on recent product which shipped. other than 2 wacky edge case bugs which are being fixed. everyone has been commanded to use it.

we're not talking about some shit tier web app shit. this is real SILICON development at 5nm at near reticle limit and now moving to 3nm chaplet design.
>>
>>108096434
>100 indians is better than 1 indian
>>
>>108096454
> I hope you enjoy being sued to hell for malfeasance on your first security vulnerability being escalated to a full-blown intrusion and data leak.
You can’t be sued for that. You can make an app that destroys the users computer. Go ahead. Nobody cares. You can just do things. There are no rules for software.
>>
File: 113777-0b21d44.jpg (76 KB, 400x363)
76 KB
76 KB JPG
>>108097368
>>108097369
SAAARRR SAARRRSS M'AM DO NOT REDEEM
https://www.youtube.com/watch?v=rBhbxVmolTE&list=RDrBhbxVmolTE&start_radio=1
https://www.youtube.com/watch?v=92ydUdqWE1g&list=RD92ydUdqWE1g&start_radio=1

Tunak Tunak Tun
Tunak Tunak Tun
Tunak Tunak Tun
ye aapaka pasandeeda gaana hai, is par naachie.
naachie.
naachie.
naachie.
>>
>>108097407
what company?
>>
>>108097466
SAAAAARRRR
SAAAAAAARRRRR
SAAAAAAAAARRR
Y U NO LET ME WASH THE FLOOR WITH TOILET WATER, SAAARRR??

सब लोग अब डांस करें
sab log ab daans karen
>karen
sab log ab daans karen
> करें

https://www.youtube.com/watch?v=VfWJqrTz40I
NO NO NO HE DOES NOT CAREEEEEEEE
HE THINKS THERE ARE NO RULES
CANCER NO CARE
CANCER DOESN'T CAREEEEEE
CANCER DOESN'T EVEN REALIZE HE'S GOING TO DIE
>>
its really funny to me that a retard with a PC can sit in is room for a weekend, fine tune a local LLM for development and make better software than some faggot who spent 4 years minoring in gender studies while earning a CS degree.
it's actually LOL
>>
>>108097407
> this is real SILICON development at 5nm at near reticle limit and now moving to 3nm chaplet design.
SAAR SAAARR SAAAARRRR SAAAAARRR SAAARRRRR
CURRY cURRY CURRY CURRY
VERY VERY VERY GOOD

>>108097493
SAR THAT'S BEING ALBEIST AND TRANSPHOBIC AND WE DO NOT ACCEPT THIS IN THE UK AND CANADA.
SAR, RESPEECT OUR WOMEN WITH PENISES, SAR. SAAAAR :D :D :D
>>
>>108097189
Just because it can code better than you doesn't mean it can code better than the rest of us.
>>
>>108097513
yeah, it does
>>
>>108097400
no
>>108097480
a made up company
>>
>>108097536
Then show an example.
>>
>>108097536
>>108097542
No it doesn't, ranjeet. An English speaker can code miles ahead of you. Where did I put my slipper and belt, this ranjeet doesn't listen to his master.
>>
>>108097513
So? The average dev is probably worse than a llm. You can't and won't code everything, mister 10x engineer.
>>
uh oh Ranjeet had a clitty leak
>>
File: india ai.jpg (145 KB, 1128x1286)
145 KB
145 KB JPG
>>108094057
No. All of these replies are
>I WORK AT X COMPANY AND I VIBE CODE ALL DAY AND MAKE... nothing

If there was any uptick of shit being made, we would have a lot more interesting sites or apps, the internet would be exploding with hundreds of useable apps
But it's not.
If anything every year the quality of the internet lowers
Apps get less intuitive
More shit breaks
Companies are absolute garbage in every aspect now
We don't even have a 4chan 2.0

Don't get gaslit into thinking things are better, trust your own eyes and notice yourself

The only real change that has happened is Indians looking at more monkeys online
>>
yeah it makes sense because nvidia hasnt shipped properly working drivers in 18 months
>>
>>108097275
I hope you get severance when you are replaced by LLM jeets.
>>
Where most people go wrong with vibe coding is that they only use one model.
I use claude code with opus to generate initial implementations, then I have codex 5.3 critique and optimize the code. When you have the strengths of two powerful models covering each others weaknesses, it tends to work out very well.
I also only vibe code when I'm making edits to something that already has a stable and sensible foundation. Making entire apps from scratch with AI isn't quite there yet.
>>
The problem is not the llm, it’s the human prompting the llm and signing off on its output. Once we can totally eliminate the meat machine and its numerous vices from the system, we will finally have good software.
>>
>>108097704
Kind of. At this point if you know how to achieve certain things in code and you can give hyper-specific instructions, it will generally follow it pretty well.
>>
>>108097286
This sounds like more work than just writing code yourself. I could see vibing something to initially get started so you don't have spend all the initial time writing boilerplate. But beyond that, it seems a lot easier to just write your own code instead of bothering with all this.
>>
>>108093825
Get ready for Microsoft levels of shit driver updates.
I wouldn't go updating any time soon.
>>
I honestly am struggling to tell what is shitpost, baitpost, botpost, realpost, truth or lie any longer.
>>
>>108097704
Not anytime soon. The training data itself is corrupted with meat shit. All of those coding conventions, best practices, architectures that make code "easier to reason about" for meat retards, it's all still there, and all of it must go.
>>
>>108094310
Thats not a problem. The problem is whether the code works or not and whether people can ship products.

If it builds and the feature you want works, ship it
>>
>>108097379
>>108097756
Most tasks can be described in a moderately detailed far more quickly and easily than they can be implemented normally. There are some tasks where it's easier to just do it than to explain it, and some tasks where I don't trust it to get involved at all. But those tasks are getting fewer in number as the models get better. I haven't properly timed myself, and I know some studies found that devs were overestimating the productivity increases.

Time aside, I definitely find it easier which is helpful for avoiding burnout. Let's say I'm working on some nestjs+react full stack webshit and there's a feature to add. There's normally quite a lot to do regardless of how simple or complex the feature is.
>decide what data needs to be stored and where
>if it's stored in postgres then I need to write a migration
>if we're using an ORM then I need to wrestle with that, if we're not then I need to write types
>write the service layer (new service or is an existing one appropriate? is this pure CRUD or are there business rules to apply? does it interact with other features in potentially tricky ways?)
>write controller methods
>write frontend components and page
>do all this while keeping the requirements and overall point of the feature in mind to make sure I'm not overengineering some shit that nobody will use
It's all quite scattered so there's lots to keep in my short term memory as I jump around.

Writing a prompt, even a long/detailed one, is easier. I can focus on the requirements while adding some pointers to how certain things should be done (based on experience with where the AI can get confused). Reviewing the plan it generates is very reassuring and allows my prompts to be shorter and higher level than they'd otherwise need to be.
>>
>>108097906
Sorry man, but to pass the code review you will have to prove to me that you understand the code :)
Ugh, I know. It works, tests pass, but me and you, we're good developers so we need to understand the code. We just have to, ok? That's what makes us good. Let's hop on a quick call? I'll invite some other colleagues too.
>>
>>108097925
show us such a prompt, please
>>
>>108097704
Yeah. It would just be better if they removed all code generated by humans from the data and made the LLM do from scratch. Like the chess ones where they get a fresh start. Human code stinks
>>
>>108097925 (Me)
I should add, you can expect generated code to be worse than handwritten code in terms of general quality. I always review and usually refactor it a bit, but unless I spend a lot of time on that, it's still going to be a mild to moderate downgrade. For personal hobby projects or open source contributions, that might not be acceptable. I take pride on those and don't want to send subpar PRs to fun FOSS projects.

For my dayjob though, I just accept it. Obviously I need to be satisfied that it's secure, correct, not horrendously slow, not a complete liability. But when it comes to clarity/structure/DRY/good comments/all that... fuck it, I'll cut some corners. My coworkers didn't seem to care much about code quality before AI and they definitely haven't started caring now, so I just match their energy. Whoever has to change it later is probably going to use AI themselves anyway. And again, this is becoming less of an issue the better the models get.

>>108097942
Sorry, no. I don't usually save them, and the few that I have a copy or screenshot of are too specific to my job.
My main suggestion for anyone getting bad results is to play around with plan mode for a while. Try to get it to write a plan that seems reasonable, something you could follow yourself. You'll get a feel for how much detail you need to give, when to give it some hints, stuff to add to AGENTS.md.
>>
i vibecode as a total nocoder and it gives me prorgrams and it's "free"



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.