[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


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


File: 27373626272.png (818 KB, 1151x661)
818 KB
818 KB PNG
>well it can't even write code that works
>well it can't even program as well as a junior developer
>well it can't even program as well as a senior developer
>well it can't even optimize this code base
>well it can't even understand this new library
>well it can't even code from this spec sheet
You are here
>well it can't even build my saas project without running out of tokens
>>
>>108191211
i like using them but youre just a gay useless faggot with no thoughts or creativity so youll do nothing even with their power, ever
top kek
>>
>>108191264
coping and seething
>>
Codejeets are so over lmao
>>
>>108191211
Nice bait
>>
>>108191211
you are the same guy that post the shitty wojak.
Why don't you go outside and take a walk ?
It would be good for your health
>>
if codemonkeys didn't want to get replaced maybe they should have picked a harder job
>>
>>108191211
In what possible scenario do LLMs serve perfection? What the fuck is that strawman image you posted, you Indian retard?
Never have the gall to ask where all this Indian hate comes from, it's you specifically.
>>
god i love vibe shitters. they're going to make so much garbage for the corporate world.
>>
The whole thing about programming is that it's supposed to be certain and perfect in a mathematical sense. AI imitates humans, which is the exact opposite.
Vibecoding combines both, which is, in short, absolute bullshit.
>>
>>108191264
Wow you got triggered hard
>>
>>108191474
>In what possible scenario do LLMs serve perfection?
In one try? Probably not for many many years until it can literally read our brains.

That said, how many super smaht human software engineers serve perfection in one go? I'll give you a hint: less than one.
>>
>>108191683
>The whole thing about programming is that it's supposed to be certain and perfect
Oh ya just like (let's see) the Linux kernel, right? Certainly certain and perfect and no bugs there...
>>
>>108191693
who is the coper and seether now
>>
>>108191766
I live in a perfect world, I like to be happy
>>
>>108191411
kek
>>
>>108191778
Still you bro
>>
File: 1771214773595880.png (13 KB, 720x738)
13 KB
13 KB PNG
>>108191211
>>
>>108191683
>it's supposed to be certain and perfect
Even Unix has bugs. Hell, even Emacs has bugs and those are written by true autists
>>
>>108193879
>if I post this it will make AI go away
bad news chud
>>
A guy in our team started working on a neural network in December. The first version was vibe in a day, including AWS infrastructure and everything, and it was done by a pm, not a programmer.
The boss was so impressed by this, that the whole company policy changed, we are now required to do 95% of the work using AI.
The NN is now literally on version 87 and it still doesn't work. We have been fixing it for months now and in an hour I will go to work and manually check every commit around the time a specific buggy behaviour was introduced. In 3 months an engineer also could have made the NN, connected it with the rest of the system and we would actually have clean code and at least 1 person would understand the code.
>>
>>108193914
That's what senior devs are for now. Your job is to fix code that retards vibe coded.
Once you burn out there till be no replacement programmers available and once enough senior devs burn out there will be no more competent programmers left and AI will be the best coder by default. Then we will all use jeet level software .
>>
>>108193914
Skill issue. Not surprised. Most swe still use AI as just auto-complete (copilot) or write functions (cursor). If any swe on your team actually took the time to learn agentic coding with cc/codex/antigravity you would easily have whatever basic neural net model up and running bug free in a day
>>
I've been a software developer for about 15 years and I don't know anyone working in tech that's been replaced by an AI. And I've known people who got fired for all sorts of things. I've seen my director fired and get escorted out of the building. I knew another guy that got put on some weird probation because he got into legal trouble involving an underage girl but for some reason they were too pussy to do anything until it was all settled. I know people who are fucking useless too and if AI can't take their jobs then it's fucked.
>>
>>108193914
When I see posts like yours, I seriously wish the whole project were public so everyone could see it and we’d know whether it’s really a case of AI being incapable.
>>
File: 1770050620645.jpg (251 KB, 782x545)
251 KB
251 KB JPG
Why can't it make the jump? I want holographic meatloaf.
>>
>>108191211
Its currently at level of junior dev with savant level skill. If the performance wouldn't nosedive after passing a certain limit in context window it would have been at a senior level
>>
>>108191683
If that were the case there would be bug free software. But thats almost impossible to have at scale be it humans or AI
>>
I put AI slop into production at work because I'm lazy. It's very rare that it one shots something. What I ask of it needs to be very simple for that to happen. You fags are shilling it too hard. It's almost like you want to exaggerate how good it is as a contrarian knee jerk reaction to normies hating on it. You're just going to get non-normies hating it with this excessive shilling. Leave some room for nuance you stupid fucking faggot.
>>
>>108194108
It's not as bad when you know what you're doing but you can't just vibe. The first few iterations of the NN were already wasted because of lacking domain knowledge, but having a good spec for that would've taken time so the gains would no longer be as impressive. You would still need a domain expert and an engineer to figure it out.
But I think there's also a technical reason it doesn't work well. The data generation scripts, AWS scripts, inference code all have to fit together and iteration cycles are slow. With just code the AI just tries 100 times until it compiles, but when you have to train a new NN for hours that approach is no longer good. There also isn't an extremely clear success criterium, the NN will never be 100% accurate so the AI can't just change the code 100 times until some unit test passes. Sure, you can look at the Huber Loss, MAE etc. of the network itself but several times it just learnt from garbage data so it learned how to predict garbage accurately.
At least currently AI is not something that can just be used by non technical people for complex projects. And if you're an SWE you still have to use your SWE skill, you just don't have to type out every line.
>>
>>108191211
I had some Claude generate unit tests for a small library project of 10 or so classes the other day.
It came up with just under 100 tests total.

Roughly 10 of those involved actual integration tests with an embedded test/dummy web server and were not asserting success status codes before attempting to parse content from the response, which would cause runtime exceptions

Roughly 60 of those didn't keep to the arrange-act-assert pattern correctly; and 30 violated the naming conventions for the tests.

20 had incomplete assertions.
5 had entire test cases missing.


>well it can't even write code that works
You are, infact, HERE.
>>
>>108194163
>limit in context window
Skill issue again. Each feature/PR should be broken up into groups of tasks which should each be run with a new context and take no more than a few messages per convo.
Op is right. Currently the only real limit is to do with the actual tool usage limits due to hardware limitations.
>>
>>108195851
But it still needs to have context of the entire code base when there are multiple dependencies and moving parts across the board.
>>
>>108191211
>Can't even write code that works
Who said that, at any point?
>>
>>108194196
>It's very rare that it one shots something
>I, on the other hand, write perfect, bug-free lines of code that run/compile flawlessly the first time I touch my keyboard
>>
>>108194351
Skill issue
>>
>>108191211
It's still the same as it ever was: only good for one-off scripts.
>>
>>108195872
Literally everyone just a couple years ago when they first came out? Do you live under a literal rock?
>>
>>108195883
skill issue
>>
>>108193914
>The boss was so impressed by this, that the whole company policy changed,
>and it was done by a pm, not a programmer
How will AI solve the problem where the dumbest Dunning Kruger retards are rewarded by society?
>>
>>108195878
buggy code you wrote and understand >>>> buggy code nobody wrote and nobody understands
>>108195851
>skill issue
>tool issue
>hardware issue
so two more weeks?
>>
>>108195987
>you write buggy code only you understand
>die
as opposed to
>AI writes buggy code that can be understood by anyone that runs the same context
>AI improves and never dies
>>
>>108195916
By becoming the ones (things?) that run society. How many times do we have to tell you Terminator is a documentary.
>>
>>108196761
There is no "the same context". Each run is different once you restart or compact. You can let AI write summaries but it's not the same.
>>
>>108196761
>AI writes buggy code that can be understood by anyone that runs the same context
If that were true people wouldn't check in buggy or vulnerable code. People don't even fucking read it anymore.

>AI improves and never dies
Until they hike the prices.
>>
>>108195881
two more weeks - right?
like they've been saying since.. what? 2022?
>>
>>108191211
My job has mandated all PRs must be entirely generated so I’m going to slopmaxx and not worry about it. Let them suffer the consequences
>>
>management insists we use AI slop generators
>it produces shitty, unreliable code
>spend all of my time arguing with a robot when I could have coded it by hand in less time


I don’t want to live on this planet anymore.
>>
>>108195881
>>108195898
If the ai is so smart why does it break if you hold it wrong?
>>
>>108198496
Its as smart as it is, you need to learn how to leverage it best
>>
>>108198499
>it’s not dumb, you just need to know what the fuck you’re doing enough that you could write it yourself instead
so it’s not replacing anything, got it
>>
>>108198499
>you need to
>noidontthinkiwill.webm
>>
>>108198516
>so it’s not replacing anything, got it
Not yet
>>
Uh oh I don't see coders on the list
>>
>>108193962
If this is a skill issue, that means working with AI still requires programming skills kek
>>
>>108200666
the only safe people in that image are hr managers who control hiring, senior executives who are at the top of the company, judges who can thwart laws, and diplomats who have a tiny bit of control over the law. everyone else will eventually be at risk.
>>
>>108193962
I've used Codex GPT-5.3 (Very-High) AI agent btw
>>
>>108200716
The hubris of man, to dare correct the word of the all-knowing robot. Pitiful.
>>
So I should stop learning new things about /g/ stuff now and give up?

It's the only thing I like to do
>>
>>108200746
No, they're just Al cultists and bunch of achizos
>>
>>108191211
>well it can't even write code that works
We are here.
>>
>>108200716
Hiring will eventually be managed by AI too
>>
>>108193934
>That's what senior devs are for now. Your job is to fix code that retards vibe coded.
This is the way. In our big corpo company, some teams have switched to org structures like "1 guy on adderall shits out massive amounts of trash using 10 claude code instances running in parallel, 3 guys shovel through this shit with claude code, 5 guys apply final touches and merge PR's". It is insanely efficient. We accidentally made our own Jira that works faster and looks better than Jira.
>>
>>108200948
Another Indian? yes...
>>
>>108200956
>We accidentally made our own Jira that works faster and looks better than Jira.
kek this is one of those things that sounds great to a nocoder. real ones know that a retard could make a ticketing system that mogs the average jira install, because every shop bogs it down with horrific workflows and plugins.
The moment the PMs got their mitts on your new Jira it would plunge deeper into hell than you can imagine.
>>
>Hey, this is amazing! The AI just one-shotted my code request!
>I just need a few modifications and it'll be perfect.
..and on the second mod it begins to break everything.
>>
I don't care, Anthropic is paying me to provide human feedback, I will basically never be out of work because the models can always improve.
>>
>>108195883
Said the tard who never interacted with an agent running the latest models.
>>
AI isn't going to automate coders, it's going to automate jobs that require being a celebrity like hollywood and the music industry.
>>
>>108191633
So nothing will change then lol
>>
>>108200716
AI doesn't have accountability built-in
>>
>>108201225
>the models can always improve
You're assuming endless cash though.
>>
>>108196863
>If that were true people wouldn't check in buggy or vulnerable code.
>implying people have never checked in buggy or vulnerable code before AI

>People don't even fucking read it anymore.
>implying people read every little bit of code in a PR before AI

You are peak Dunning Kruger and you don't even realize it
>>
>>108198496
skill issue but actually these days it will flat out tell you when you're obviously wrong. try to convince it that division by 0 is possible for instance
>>
>>108201448
>>implying people have never checked in buggy or vulnerable code before AI
>implying there's no difference in bug/vulnerability rates before and after AI
>implying vibe-coded software doesn't get opened *regularly* now

You sound autistic, and I don't debate autists. I put them in cages, where they belong.
>>
How many popular steam games has AI made??
>>
>>108201468
>implying most software hasn't been jeetified and riddled with bugs before AI
Not fooling anyone, Jeet. You're getting replaced.
>>
>>108201482
>most software
I am so lucky to appear to have avoided that fate.
Enjoy the cage.
>>
>>108200956
>This is the way.
Yeah, the way to make total dogshit and also make sure that the company can no longer even pretend to make a competent product in 10 years.
>>
File: pepe-the-frog-laugh.gif (107 KB, 498x498)
107 KB
107 KB GIF
>>108195881
It's too much fucking work for something you can do yourself if you know what you are doing

>but it makes ME feel smrat!
lol
>>
I already have my own vibe codade replacement for photoshop and vlc
>>
I don't know why some programmers are trying to resist coding agents. it literally saves so much time.
>it produces slop
not your problem. the code belongs to the company - you should just switch job before the slop implodes
>i like writing code by hand
do it for your personal project. nobody is stopping you.
>>
>>108202029
Ok, but why? Both of these can be had for free with very little effort
>>
>>108198499
Already know how to leverage it best:
not leveraging it at all.
>>
>>108202029
Is it on Loonix? If it's actually good publish it pls.
>>
>>108191211
sure. but can it do something worthwhile?
>>
It's retarded to be proud of your prompting skills (even if we accept such a thing exist, just for the sake of argument)
If those tools really get as good as AI fans predict, using them will become easier and easier. Also a lot of configs are already online, and generally prompting is much easier than real programming. If LLMs become really good, people who are not currently using them won't be "left behind", they will just clone some config from GitHub and learn the remaining details over one weekend.
>>
>>108191211
Your fallacy is thinking that because it can do it sometimes it can do it every time while in reality it can do these things less often than it can't.
>>
>>108204208
Stay behind luddite
>>
>>108201395
Bro, have you seen Anthropic's revenue growth? Did you know most applications in the world are just forms and database writes/reads, which are the perfect candidates for AI gen code? I get the skepticism, but sometimes all this feels like downplaying the internet in the 90s. This shit is far from perfect and sucks at a lot of things etc etc but it will change the world that's not even a question anymore.
>>
>>108205019
I honestly don't understand why a lot of stuff wasn't automated way earlier. There is so much software that basically does the same thing, everyone wants their own special snowflake code.
>>
>>108191211
>AI code more popular than ever
>almost all big software that uses AI runs worse and has more security breaches than ever before

lol, lmao
>>
>>108191211
It keeps making progress and if this continues we'll all be replaced.
>>
*throws a pipe bomb at your window*
>>
>>108191211
who is in the best position to apply llms to real programming tasks?
>a, the programmers already doing so
>b, someone who doesn't know what a compiler is

the gatling gun was invented with the goal of reducing army sizes and battlefield deaths
afterward, there was:
>the american civil war, 3 million combatants
>world war 1, 70 million combatants
>world war 2, 100 million combatants
who was in the best position to apply machine guns to war? existing trained soldiers

the denial is with nocoders who think expansionist companies won't expand to capitalise on what will become possible with future llms, that codebases will remain the same size and solve parochial problems
>>
>>108191211
LLMs have unironically gotten worse at OOD coding tasks (ones not directly represented in the training data). o3 and Gemini 2.5 Pro were peak at that. But GPT-5 and Gemini 3 model families are absolutely fucking worthless. I have to use Kimi now for things like making minecraft plugins, etc.
Since all the western corporate models are useless benchmaxxed garbage now.
>>
>>108200948
hr kisses C suite ass if not outright sucking cock.
the nepo circle-jerk at the top will go on for a while before AI affects it. Hiring their son or mistress and having them copy paste ai responses probably makes it easier.
>>
>>108191211
You are here
>well it can't even write code that works
>well it can't even program as well as a junior developer
>well it can't even program as well as a senior developer
>well it can't even optimize this code base
>well it can't even understand this new library
>well it can't even code from this spec sheet
>well it can't even build my saas project without running out of tokens
If AI was good, AI companies wouldn't be giving you tools built on Electron.
https://www.dbreunig.com/2026/02/21/why-is-claude-an-electron-app.html
It can't code, it's copy paste from stackoverflow/github/etc, with more steps. It's "actually indians" an automated jeet.
>>
>>108197985
This is the way.
>>
>>108198377
Who gives a shit. Just push that slop to production, lol.
>>
>>108207999
>just actively burn everything around you
WW3 would be quicker at this point.



[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.