Hello programmers! I’ve been making a living for a couple of decades figuring out stuff for people and coding whatever will make them happy. It’s fun. I’ve grown to prefer the people to the coding but I still get a kick out of the whole thing.
The problem is: at the job we’re being asked to pivot to vibe coding. At this point the order is to use it for all code, read the slop carefully, argue with the machine for way too long and only then open that PR. Massive productivity gains are expected. I’m appalled from an ethical, philosophical and professional standpoint. We’re a non-profit ngo ffs. My colleagues are fine with it. If I get even slightly critical the group goes awkwardly silent and I get comments on how all the programmers they know from other places don’t code anymore.
If this is how it goes from now on, I’m out. Is it really though? Are all businesses switching to vibe coding? Have some of you started switching career? I think I want to go to meatspace. Hand-made coding could make a cute hobby.
No, vibe coding is what happens when a new technology spawns, and everyone is in a gold rush to find out what they can use it for.
LLM assisted software development is probably here to stay though. I’m also learning the ropes of this agentic stuff, where it works and where it falls flat as I always try to fully understand and take ownership of the code I put in my commits. At the end of the day, it needs to look like I coded it.
On an aside, I’m hopeful we can figure out an ethical way to do model training that respects licensing, such as only training on MIT licensed code. Such a rabbit hole the ethics are.
It’s not going away.
Everyone who complains about the quality of the code it produces is overlooking one important fact: 90% of the code developed by the software engineering profession is utter dogshit, pasted together by people whose understanding just about extends to typing in stackoverflow.com by banging their head against they keyboard.
LLMs are not replacing artisinal hand-crafted code developed by a skilled engineer from a top-10 university, it’s replacing the utter shite produced by the people with an NVQ from Timbuctoo Community College that actually powers the vast majority of websites and enterprise software.
And it’s fucking good at it.
Look, when I started my career I used to hand-write assembly code. I used to personally review the output of the Keil C51 compiler because its optimiser was objectively dogshit. Nobody is doing that any more - partly because the compilers are better than programmers at writing assembler these days, and partly because we have so much CPU and memory we don’t know what to do with, so who cares if they aren’t. Both apply to LLMs, and we will soom be at a stage where reviewing - or even caring - about the quality of LLM produced code is as anachronistic as thinking you need the check the compiler’s output. (TBF, Clang/LLVM for AVR still produces dogshit assembler, but that’s by the by.)
Just as assembler did before, high-level code is going to become just another intermediate representation that isn’t really important; what will matter are the prompts and context you give to the LLMs. And there’s still a hell of a lot of engineering skills that need to go into that.
Software Engineers - the real ones - have nothing to fear from LLMs, and in many ways it’s much more satisfying working with an LLM than a team of barely skilled “coders”. The barely skilled coders who wouldn’t know malloc from a semaphore though? They’re fucked.
God I have þoughts about your post, but I’m not typing all of it out on my stupid phone, so maybe if I remember when I’m at a computer þis comment will serve as a placeholder.
Well said!
Just as assembler did before, high-level code is going to become just another intermediate representation that isn’t really important; what will matter are the prompts and context you give to the LLMs.
Ah the old “LLMs are just a new layer of abstraction” argument. No they aren’t, and won’t be any time soon.
You mean that “write this service make no mistakes” executed with a LLM-as-a-service where the underlying model updates every week and the model itself is deprecated in 8 months isn’t a good idea for a source code representation?
Nah, it definitely is. It’s called job security.
On the first half, that’s not how it’s done; the input is a much more detailed description of design and architecture coupled with TDD.
On the second half - I happen to agree, that’s why I think the OpenAI/Anthropic business models are doomed to fail. But that’s where locally deployed models become key - predictable output.
They don’t need to be better than the best programmer in the world to change how 90% of code is written, they just need to be better than a junior dev from Timbuctoo. And feasably-deployable open models crossed that bar around Q1 this year and are only getting better (Qwen3.8-Flash-Next is another huge leap forward.) We may not be quite there yet, but the destination is now inevitable; adapt or find a new career.
On the first half, that’s not how it’s done; the input is a much more detailed description of design and architecture coupled with TDD.
Yes, I have these coworkers too. I’ve seen what they look like, and it doesn’t change my point.
But that’s where locally deployed models become key - predictable output.
Local controlled models help, yes. That isn’t what business do, though. And even for local models, the models are still stochastic as fuck. Maybe that changes in the future, in which case I’d reconsider that view (though natural language is still a terrible interface for telling a computer precisely what to do).
We may not be quite there yet, but the destination is now inevitable; adapt or find a new career.
How many years has it been now? Mythos was the end of the world, Fable had to be locked down by the US government, Astra is going to cause humanity to go extinct, and yet here we still are today with vibecoded, mass-produced slop.
it’s not gonna last. if you tune out of the hype and actually look at some hard data, vibecoding stops looking so hot: https://codemanship.wordpress.com/2026/08/12/ai-software-development-what-does-the-data-say/
the current token costs are unsustainable also. the only ones making bank right now are the shovel sellers: https://isaiprofitable.com/
personally i’m working on hand-rewriting several projects vibecoded out by a non-dev in a small shop. not only is a lot of this straight up broken nonsense, the amount of tech debt is staggering. previous job wasn’t better, i had to spend more time fixing slop from former coworkers than actually introducing new features. keep your skills sharp and persevere. we’ll get through this.
tl;dr: this, too, shall pass. but holy shit.
It’s absolutely going to last. IMO until Astra the models were not good enough to really be worth bothering with, but we’ve reached a tipping point.
And while I still do read the code they make and then fix things (or tell them to do things differently), it’s still waaaay quicker than doing it “by hand”.
Sorry but if you’re expecting this to be a temporary thing you’re going to be very disappointed.
I agree with you except for the Astra bit. The differences in day to day are marginal at best (compared to both Sol and Anthropic equivalents) and the paradigm shift came long before it.
I dunno about Sol but Anthropic models at least are waaay slower (at least x2) and also write that really annoying AI prose - it’s not this, it’s that, and so on. Astra doesn’t do that, or at least not enough to be annoying.
Sol is pretty fast from my experience, but still does the same thing that everything before it did: generate plausible code that I need to correct (and often rewrite) afterwards.
I haven’t used Astra yet, but I don’t really expect it to change that. Once it becomes available at work, I’ll give it a shot of course, but my expectations aren’t very high.
It’s copium, which is understandable. I think it’s really important for people who want to have a reliable career in the industry 5+ years from now, to learn harness coding even if begrudgingly.
You know the current state of IT security? It’s bad.
Aand here comes AI. It will not make it better, even though it finds some issues faster.If vibe coding remains the norm and is actually sustainable, there won’t be a reliable career. You’ll be weavers being put out of work by the autoloom. You’re worthless to them. What can you do that someone with six weeks (or less) of training can’t?
What can you do that someone with six weeks (or less) of training can’t?
AI seems to be killing opportunities for entry level engineers. Hiring and interviewing processes I’ve been involved with or observed in the last ~2 years have greatly favored experienced engineers, just less of them. I’m gonna do my damndest to adapt, network, and remain in that bracket.
just less of them
That’s not a trivial problem. There will still be some programmers, but everyone will be competing for a shrinking number of jobs.
That’s not a future with a stable career. That’s a future where you’re going to have to fight for the rest of your life to keep a job.
Unless the LLM hype craze dies down and they realize “oh actually we can’t replace people yet” and they start hiring again.
IMO until Astra the models were not good enough to really be worth bothering with, but we’ve reached a tipping point.
People said the same about older models.
Yep. Every. Single. One.
Everyone downvoting you is doing it while sipping their cups of copium.
If they continue to ignore, then they’re gonna find out just how right you are. Real quick.
But being able to ship small, mildly broken, internal tools for my non-technical coworkers quickly is actually amazing.
Nobody cares if the internal tool code is nice, but when it shaves 50% of the work off of sales the sales team gets excited. So if I vibe up a shitty UI it’s still a net gain.
I’ve probably automated away like 60% of the ad hoc work I used to get into internal tools, and that has helped a ton with workload and budget.
Careful. You know how it goes with temporary workarounds? Suddenly, you find yourself maintaining slightly broken load-bearing slop.
That’s a fair concern, usually I only vibe the UI.
I’m not a front end guy, so as long as I know the backends is good I’m fine with it.
Ok, UI coding is messy either way.
Yeah, it’s only the user’s experience fucked over when the UI is sloped together.
Although I agree with you…
That’s how the vast majority of software has been built for the past couple of decades. LLM’s aren’t different in that regard, just faster.
That’s exactly how a webpage that displays an excel spreadsheet nicely in different tables, turned into a gigantic monster of a project that nobody fully understand at my workplace.
Well, kinda, because that was BEFORE ai. I can’t even imagine what mockery of God himself they would have done with AI in their hands.
I am a technician in the physics department for a major university. There are quick and dirty hardware hacks that I threw together 2 decades ago that are still in daily use. Quickly made mildly broken tools have a habit of lingering.
keep your skills sharp and persevere. we’ll get through this.
We’ll be the next generation that writes malware no one can stop
Upvoted because I want it to be true
I’m a systems architect who uses agentic coding to get around some executive dysfunction challenges with writing my own code. I learned about the dysfunction late in life; maybe meds will eventually get me to a point where I can consistently work on a personal code project without a looming deadline or social contract in the loop. For now I am part of the problem. Hi!
I look at agentic coding as a firearms problem:
- it lowers the barrier to execution: anyone with sloppy aim can point and pull the trigger, and do a lot of harm in the process
- both perform better in the hands of people with discipline and experience, but that doesn’t change the fact that LLMs and guns have massively proliferated and any idiot can get their hands on one (or that shitty people benefit from mass producing them)
- the industrial complex producing guns profits from murder, and the industrial complex producing datacenters profit from murder of the environment and attempts to eliminate jobs
- the same countries who treat gun control as an anathema tend to view AI control as an anathema
- analogies suck, they’re never perfect but they get the conversation started
That said, most of the people who hate AI also don’t know a damn thing about how they’re used correctly. It’s also not their fault. Sure, they’re under no obligation to be an expert about the thing that’s harming them. But there’s a bigger reason: the companies pushing this shit benefit from a world stage where people use them wastefully and inefficiently.
The short list of things most people don’t get:
Spoiler'd because most people also won't give a fuck.
- A LLM is like a ten year old with a doctorate who wants to impress you. It can answer random trivia off the top of its head, but it’s going to make shit up when it doesn’t know the answer.
- Never trust a LLM answer that was produced without a web search or web fetch. Pretend you’re that obnoxious algebra teacher who rolled their eyes at your calculator and told you to show the work during exams. Work for a LLM is tool calls: web searches and webpage fetches. Tell it to research documentation and best practices before you start. If you don’t, the answer you get will be dated by 1-2 years in the best case, and a confidently wrong hallucination at worst. (unfortunately you’re still at risk of being fucked because frontier labs use webpage summarization in these tool calls, and that summarization adds another vector for hallucinated shit; lovely)
- The baseline skills of the operator are more important than anyone gives credit. Your results scale to your own knowledge level and your use of the LLM to extend into adjacent subject matter. Employers don’t want to hear this: they want employees to go away. Frontier LLM labs don’t want employers to hear this: they need employers to buy into the delusion of eliminating their work force, and also financially benefit from people using the tools wastefully. (token burn = profit = the heat death of the universe as an afterthought, but that’s the next generation’s problem) Casual LLM critics don’t want to hear this: confidently incorrect echo chambers are the comfort food of anyone with an axe to grind about what they hate.
- Consistent results require controlling the probability space so that your pipeline swats the hand of the LLM when it’s lazy or generating slop. For coding, that means using languages or tooling that prevent major classes of problems to begin with (Rust, Go, Typescript) or toolchains that bitch about sloppy code (pytest-ruff + pytest-ty for Python). It means extensive unit tests that cover expressions of functionality, and knowledge of how to prompt out bad behaviors. (don’t rewrite tests to conform to “preexisting bugs”, always align tests to higher standards, etc.)
I could add another ten bullet points here, but I’m not here to prove that LLMs are capable of being useful. I’m here to demonstrate that it requires knowledge of a problem domain to get the best results. That means you can’t get rid of employees familiar with the knowledge domain. You need senior employees to get the best results, and you need junior employees so that they can eventually become senior employees. Same as it ever was. But because we’ve got useless fucking politicians who are too busy lining their own pockets instead of, I don’t know, regulating a worldwide memory shortage caused by a handful of companies, the whole thing is cannibalization of human society all the way down.
We could have scaled this to operate in the consumer space at a fraction of the energy costs (Apple SOC style versus energy inefficient hyperscaling) and used it as something to generate value by augmenting the capabilities of individual workers instead of trying to replace them wholesale. It only makes sense that it should have benefitted everyone considering this tech was built on shoving libraries worth of human achievement into a training pipeline (often without permission), right?
Nah. They profit more if we don’t own our own compute and everyone is encouraged to burn energy on treating the LLMs like all-knowing oracles instead of research tooling. They control the access to creations built on stolen knowledge. But hey, if everything else gets burned down in the process at least they come out on the other side rich and bailed out.
I am working in a very small company as a software dev and write everything by hand. I wrote some weeks ago in a different thread here that I tried all of the LLM powered coding tools and found all of them subpar.
If this company goes downhill or I lose the job otherwise, I will work something entirely different.
My very arrogant point of view is that these “vibecoding” tools are being celebrated by people who were making shit software before and now make software that is slightly less shit in what they perceive is the same amount of time, but I suspect it takes them actually longer. Meanwhile, they don’t grow their programming skills at all.
people who can suddenly “”“make”“” things like minecraft mods are also a part of those celebrating, though i assume they’re a lot less vocal about it
and god i wish that was all it was used for
From what I can observe, yeah, pretty much the entire industry has shifted to LLM code generation. I’m sure (or at least I hope) there are still some shops that go out of their way to keep working like we did, you know, months ago. But beyond that… I’m not even sure the general public understands how quickly and how fully LLM codegen has become the norm. It’s grim.
I write code. There is skill to being able to read documentation. I love tools like predictive IDE’s. Search engines augmented my ability to code with sites like stackoverflow. LLM’s are a continuation of that trend. Often the output is broad strokes that I need to refine. At my workplace, humans are required even with the tool improvements.
I’m being migrating all my team code from SAS to Python from the past year or so, slowly but firmly. Another team went full on on vibe coding their SAS to Python and now I’m being told that I’m being move to their team to refractor all that vibe code because the results are incorrect and nobody understands what the vibe code is doing. I checked up quickly to know what I’m going to find and the Claude Code just translate the SAS code as is, instead of adapting the code as python practices (for example, to get the last day of the month the code goes to the first day of the next month, but first it check it December to go to the next year, and then subtracts a day, instead of using relativedelta).
I’m still designing and writing all code by hand. I don’t know that any of my coworkers are and know that many definitely are not. Our company does have copilot reviews built into CI and, at least for now, requires human review as well.
I read a retrospective from another team yesterday. They were preparing a demo for leadership and apparently it didn’t go so well. There were many reasons like last minute changes in requirements but one line was interesting for this discussion: The use of the LLM agents caused confusion about who is fixing which bug and taking on which tasks. The agents act like they’re alone in the codebase and keep adding and changing code broadly whenever they see an opportunity for improvement. And that often without specifically being prompted to do so. Now if you have multiple team members collaborating on a project with each their agents, chaos ensues because the agents are not coordinating and the diffs/commits become insane and the team doesn’t fully understand them.
This actually results in a necessity to test more and longer because nobody knows which bugs have really been fixed now and rapidly preparing for a release or a demo becomes more difficult than before.
In my opinion an LLM can be very useful to rapidly prototype something. Especially with an API or language you’re not familiar with. But for Production it’s still a huge risk to push code you don’t understand.
Those engineers are bad
I don’t think it is like that everywhere though it is like that far too many places. I’m hopeful eventually people will realize it is unsustainable and reach an equilibrium where the usage will be less harmful, though we can’t say it’ll happen for sure or how long it will take if it does.
The fact everyone is listening to people saying it’s rocket fuel that aren’t the people who actually held the cart together in the before times is maddening, and it is equally maddening that the best devs have extra work mitigating the damage while it is essentially invisible and demoralizing work.
It is heavily normalized in industry now.
I quit a job that added AI use to performance reviews super early. That was pure bureaucratic cancer and most of my good coworkers burned out and left. I don’t think I’ll ever work for a big corp again, I actually like my work so much more now.
At work I’m expected to use it, our clients want it, my boss likes that he can feel technical when writing tickets (they’re mostly wrong though, and the tickets solution instead of report and request). It genuinely can do some useful tasks that make sense to use it for.
There are real pros and cons, but there’s no putting this back in the bottle. Philosophically I do hate it and everything it stands for, but practically it has allowed me to do a lot that I couldn’t (and wouldn’t) before.
The only good solution I’ve found is that you have to get your hands dirty and stay in the weeds. You cannot let it run off and do what it wants, you have to drive and then tag it in. When you vibe code something you’ll find yourself in three weeks resolving an issue and going “wait, it does what?!”, and of course, arguing with the gibberish Claude or GPT claims is English but is actually the first stage of model collapse as the models reinforcement train on ever more distilled data. Plus, skill atrophy is real. Decide which skills to allow to atrophy carefully.
AI code review is good and useful. I used to use it as a first wave for obvious stuff, but it is actually decent and high trust now. Saves me some time on that.
It’s good at finding bugs and security issues, and one thing I actually really like is having it model a system as a state machine and enumerate all the possible states to find bugs — especially concurrency bugs. I don’t work in a provable language, so this is hugely helpful.
I would be very hesitant to take a role where they push vibe coding or “agentic fleets” or loops or whatever. If you’re babysitting 6 Claude code sessions you aren’t working and you’re not in the loop on any of them. That will accumulate as tech debt and bog you down over time. It’ll suck your soul dry. It’s miserable.
Reading the slop and reviewing it “software factory” style is a very bad sign to me.
Some code is throwaway or disposable, vibing that makes sense to me. But the core work is more important than ever. Maybe the disposable code is the new part, I look at it like installing an IKEA kitchen, it’s slop, but it looks nice, the important part is the foundation and level flooring that enables the slop to stand up.
I actually like my work so much more now.
You hiring any managers? 😅
arguing with the gibberish Claude or GPT claims is English
You’re right to push back on that – I’ve been using Claudish the whole time and this actually changes everything. If we use English instead, each one of these comments could be a single, concise line. One thing worth knowing: This is not a minor boost to understanding, but a complete game changer. Understanding my bullshit has been gated behind speaking LLMese.
Nah, really, I’ve been able to get fairly decent code out of the newer models since I work with a very popular language. Not one-shotted of course, but after some change requests or manual changes. But the language these models speak (including the Chinese ones, it seems) is infuriating and they seem to insist on tons of super long comments in places where the code itself feels self-explanatory. My philosophy is to write “why” comments, not “how” comments, but not a single model seems to agree with me.
I look at it like installing an IKEA kitchen, it’s slop, but it looks nice
Are IKEA kitchens really that bad? I’ve been looking to remodel my ancient kitchen and it comes to 5-6k EUR, appliances included (minus the fridge, I upgraded that a while ago), from IKEA. If I ask a carpenter to do a proper one, that’s going to be more like 10-20k and I don’t wanna spend that kinda money lol
IKEA kitchens are great?! We’ve had ours for a long time now. And even moved once and took the kitchen with us. It’s easier to assemble than their competition, the surface of my cupboards has higher quality than some other kitchens I’ve seen. There’s no silly salesman in the process who tells you it’s 17,000€ but you get a special discount and only need to pay 10,000€ if you order it today… You can just grab additional cabinets 12 years later…
I didn’t buy the appliances at IKEA, though. And I found nicer alternatives for the sink, plumbing, countertop as well. But the rest is from IKEA and I’m A-okay. Also like my IKEA wardrobe.
Only downside is if you don’t assemble it yourself. I had a look at their price table for assembling it, and that service was a bit expensive, I think. At least where I live.
And even moved once and took the kitchen with us.
For anyone confused, this is common in some parts of the world. Taking your entire kitchen with you to a new home is a thing.
And yeah, appliances aside, IKEA has great countertops, cabinets, and so on.
Yes. I mean it doesn’t have to be like that. You’re also allowed to (try to) sell the kitchen to the next person. But the landlords usually don’t want to be bothered with the upkeep, fix appliances or get insurance in case the plumbing floods your stuff. So they regularly just provide people with the walls, heating… And since it’s a bit difficult to move the loo, the bathroom will be okay and contain a toilet, shower and sink including faucets 😅
And turns out I misspoke. I meant I got a worktop at some other place, not the entire countertop.
I think IKEA has different price points
The ones installed when I moved in were those cheap particle board counter tops and cabinets, so when my dishwasher leaked everything expanded and cracked and can never be fixed. A real wood or high quality MDF would have been okay with a one time water event (if dried properly)
It stood up for years, worked fine, wasn’t awful. Just was cheap and not resilient.
The problem with cast coding a bunch of vibe coded apps is you have to maintain the vibe coded apps.
As a developer of over two decades, ive seen ai aka llms good at:
- Creating small MVPs to prove something is possible as software
- Stealing and using other peoples code to make said apps.
- Quick very small directed scripts.
- Making up funny random shit. Such as aidungeon.
- Using up ram
And its mediocre at a lot of things.
And straight up bad at other things such as cooking and instructions that need to be followed to the letter.
If you work for a company that prioritizes profit over quality - yes.
I no longer work in government IT, but from what I have heard from old colleagues, quality is still prioritized there and they are not forced to vibe code.
I don’t think so . Ideally, it will provide some increase of productivity if it will allow easily search for some information quickly ie faster and simply than googling or stackoverflow. But as for using natural language to write software here is good answer why it will not happen : https://www.cs.utexas.edu/~EWD/transcriptions/EWD06xx/EWD667.html
