Step 1: Some upstarts create a new way of doing something. It’s clunky and unrefined.
Step 2: "Experts" and senior folks in the field dismiss it as a "toy." It doesn't follow their established rules or best practices and seems amateurish. They wouldn't recommend it to anyone serious.
Step 3: The "toy" gets adopted by a small group of outsiders or newcomers who aren't burdened by the "right way" of doing things. They play with it, improve it, and find new applications for it.
Step 4: The "toy" becomes so effective and widespread that it becomes the new standard. The original experts are left looking out of touch, their deep knowledge now irrelevant to the new way of doing things.
We're at step 2, bordering on 3.
* Executives at Nokia and BlackBerry saw the first iPhone, with its lack of a physical keyboard, as an impractical toy for media consumption, not a serious work device.
* Professional photographers viewed the first low-resolution digital cameras as flimsy gadgets, only for them to completely decimate the film industry.
The problem is not that the experts have dismissed vibe coding without even trying it.
The problem is that the experts have tried it and found that vibe coding doesn't actually work at scale that they need.
Will it ever? Perhaps, but I'd argue a near-AGI level of intelligence would need to be achieved first. When that happens, we have bigger problems (and/or opportunities?) than a few programmers losing their jobs.
Yes, you can't vibe code a full-ass enterprise level product fit to sell as SaaS for a Fortune 500 company.
But what you can vibe code is a bunch of small _bespoke_ tools that fit your exact use-case. Stuff you could buy (or rent) as SaaS, but they'd have 80% too many features and a constant monthly cost.
Just last week I converted a bunch of docker compose -based containers to run on Opentofu, took me maybe an hour or two with Claude Code while watching TV.
Would've been a week easy if I went for the artisanal route of reading Terraform documentation and digging through forums and Stack Overflow.
Not a world-changing thing, I could've lived with the compose setup, but now the whole system is declarative. I set the state and the state is applied, no drift.
So called "Experts" dismissed perpetual motion machines. PFFFFT I'd say I'm at step 4 with my PMM and I recommend you get on board, I can sell you plans for a cool 10K rupees.
Not OP, but there are many things that I know don't work without trying them. That's not a contradiction. It may or may not be true but it's not a contradiction by itself. You can know reasonable well that something doesn't work by looking at other people who have tried it (sometimes even better if those people are experts and you are not).
Cold fusion didn't work, until it did. 3D TVs didn't work, until they did. Flying cars didn't work, until they did. Satellite phones didn't work, until they did.
I feel like there's a bit of survivorship bias here. Not saying that vibe coding is completely useless, but you still have to review the code it produces in order to make the most of it really. It isn't completely autonomous (yet) to the point where it can scale to any of the examples you mentioned imo.
Naming two successful technologies says nothing. Your steps don't follow inevitably from one to another and there's a vast universe of failed technologies and stillborn revolutions that prove it. If you want to persuade anyone that vibe coding will ultimately replace traditional coding you need to give reasons you think it belongs with digital cameras and streaming and not with lisp machines, ekranoplans and every attempt at "no code" in history. Otherwise you're just making noise.
People are so invested in this bubble they have to believe it will not burst at all costs.
If this works as well as the evangelists say, they'll all soon be millionaires, right? Everyone is 10x more productive!
I'll be waiting... seated.
The good part about this is that if it does indeed get as good as they claim, anyone will be able to get the same benefits. If it doesn't, I'll be around to fix all the garbage code vomited out of LLMs.
> Professional photographers viewed the first low-resolution digital cameras as flimsy gadgets, only for them to completely decimate the film industry.
You oversimplify everything, but this is completely wrong. 320P cameras were useless to pro photographers, but as soon as the capabilities got near to what film could provide, they eagerly switched. Kodak was destroyed, but there wasn't much of a place to pivot to. "Masters of a complex, chemical domain" to "Yet another camera company" isn't a real pivot.
I'm bullish on AI, and think "vibe" coding was/is a cool experiment so don't agree with the premise that it is in the worst idea, but strongly disagree with your simplistic take on how tech is adopted. There are countless ideas that "experts" were right to ignore, ideas sitting in the trash can of history.
I saw these arguments in the crypto scene, and they all sounded great. But I also use crypto for payments (albeit rarely)..
Now this, and I also vibe code but I'm not convinced it will change too much for the coding profession. Will probably make it suck a bit more in relation to management and juniors who will make more laziness mistakes quicker.
Considering how many developers still don't write tests, pair program, or do CI and CD (shipping multiple times a day) – all things Dave argues for – I don't think it is fair to dismiss him as the establishment or incumbent.
Haha I'm not that dense. I'll give you counter examples. Bitcoin (supposed to revolutionize the world still waiting), Apple Vision Pro (supposed to change how we work, wear it on streets and airplanes, still waiting..)
Right, so why do you think vibe coding is digital cameras rather than bitcoin? Without that you haven't said anything, just made a prediction about the future without giving any reason to think it's correct.
Nobody's disputing that. The video creator himself has videos about AI coding techniques. It's "vibe coding" - using LLMs to generate code without reviewing or understanding it - that is at issue.
If you're building "production ready" stuff with logins for Other People and god forbid taking payments, yes you definitely must understand what the code does.
But when I'm building a tool that tags random meme reaction .webm files and finds duplicates from them, I couldn't care less what the quality is or if I understand what it does.
I can easily run it, observe if it does what I want and Vibe harder if it doesn't.
What have you based this model on, we seem to constantly make such broad statements of ontological truth without backing it up with any sort of rigour. Just because you can create a model that seems to fit some particular empirical truth doesn't mean that it represents some broader truth.
That's is not a fair representation of the situation.
AI as a tool for software development is here and will continue to gain traction. Everyone recognises this.
But what is touted as "vibe coding" now is hype at best and misguided.
So I would instead argue that this is not the "innovator's dilemma" at all, and in fact AI is making headways everywhere, but instead this a classic hype cycle. Currently "vibe coding" might be peak hype, it will bring disillusionment when reality catches up, and so on as we go down the cycle.
>Executives at Nokia and BlackBerry saw the first iPhone, with its lack of a physical keyboard, as an impractical toy for media consumption
This also is how most people use their smart phone today. They were right. Perhaps they just didn't realize that people wanted a on-the-go impractical toy for media consumption.
RIM was obviously attached to the keyboard because that was a defining feature of the Blackberry. They thought (or didn't want to see) that people would not want a touchscreen for typing their emails.
But Nokia had released a touchscreen phone even before the iPhone and their first iPhone-like smartphone was in 2008 so only a year after the first iPhone. What they stuck with for too long is Symbian, which was blown out of the water by the iPhoneOS and then its apps ecosystem. In contrast, Samsung quickly ditched Symbian what whatever else they were using to adopt Android.
> Executives at Nokia and BlackBerry saw the first iPhone, with its lack of a physical keyboard, as an impractical toy for media consumption, not a serious work device.
You say that like subsequent events overtook them. But I see this complaint on HN pretty frequently. I agree with it. Nobody does their work on a phone. The phones go out of their way to make that difficult.
This whole thread is giving blockchain in 2015 vibes. People were using all sorts of quotes and anecdotes to tell skeptics why they were wrong and in 10 years the entire financial system will be running on blockchain. A certain amount of skepticism and cautious optimism is healthy.
Also, people seem to be missing that "AI Assisted" coding and "Vibe Coding" are not the same thing.
Personally I think the issue with vibe coding is two fold:
1. It is not good at solving problems that are uncommon.
2. It is not deterministic.
Yes, AI can do quality control and testing now. But anyone who has done TDD can tell you that just the mere presence of tests does not itself mean the code is effective or solving the right problem.
Is it getting better? Yes. Do I trust any vibe coded apps built by people who don't know actual code and are treating it like a black box? Absolutely not.
And I say that as someone who has tried pretty much every IDE out there and uses AI assisted coding (on "agent" mode) heavily every single day.
I'm not a programmer, I work in finance, but I've read half of a book called "Python Crash Course".
I've been trying to improve my productivity recently, so I vibe coded some scripts that help me record and analyse my time. I understand the code at a high level, well, maybe 80% of it anyway.
This debate doesn't mean anything to me, I'm just going to keep vibe coding.
I have been a fullstack java developer for 10+ years and In my free time I also vibe code some python, bash or html/css frontends. But it's mostly just very basic stuff with limited scope.
It's fun having a backend, writing a few lines and after some minutes of waiting you have a fully working and decent looking frontend website available.
I suggest reading about software architecture, and reading A LOT of high quality human generated code. LLM coding tools work, when the person using it has a frame of reference for quality to compare against.
i have 20+ years coding experience, but i am a teacher not a programmer. i never coded ts. so i can easily review and fix code, but i dont know functions and parameters.
3 days ago i started modifying a kanban editor that was available for vscode. i wanted to have it compatible with an obsidian markdown kanban format i was using. but obsidian is to slow for me.
after 3 days it's not only working, but as far as i can tell it's a way better kanban editor then most of the ones i tried on vscode's extensions. of course i have specific needs (no dates, deadlines, priority), but a nicely tightly layouted interface with fast editing possibilities.
i would not have gotten this far in 2 weeks without claude code. i know i havent reviewed most of the code, except saving and loading. so it will likely look bad.
4 days into vibe coding a POC in a framework and language I don’t know, but with 40 years of coding experience: It’s amazing!
The scenario is perfect, a use case that is not currently supported but may well make sense. It’s basically sketching out an idea to let business evaluate its market viability, and to gather further end-user input.
Will the code reach production? It just might, but it at least needs review and refactoring by a developer seasoned in the framework. They might even want to rebuild it, and then they have the yard stick which to measure their output. And if they need a specification, it can be generated from the code in which ever specification format required by their processes.
The key here is that I’ve been able to iterate on the POC many times in a short time. The idea sketch has been refined, necessary details added, while others removed. Functionality swapped in and out while testing different approaches.
Right now vibe coding in this way requires substantial experience in software development to frame the problems and solutions to the AI. Without my understanding of the domain (both the software domain and the actual domain) vibe coding the POC would not have succeed.
My greatest concern is that it looks and works too good and thus will be kept as is even in production. As the old adage says: There are no temporary solutions, just more or less permanent solutions. A temporary solution that works is a permanent solution.
That is what I’m alluding to. But also, I work in a security aware industry, so those things will be vetted. And my own experience make me address some of the potential concerns already in the POC.
The real divide going forward will be between vibe coding with experience across domains vs vibe coding without IMHO.
The whole point of vibe coding is that you don't review or attempt to understand the code. If you do, you're a developer using AI as a tool, which the video is not arguing against - indeed, he links to another video he's made explaining one way of doing that. You could call that "vibe coding with experience" but it just devalues the term, I think, and certainly misses the point in this particular case.
I don't really have that much of an issue with vibe coding as an appropriate tool in experienced hands. I think the worst ideas in 2025 are probably related to IT execs pushing AI in the wrong ways, or people espousing vibe coding as some sort of software development panacea.
It's bad software practice and insecure sure but those are not things people notice and the tech industry has historically been terrible at them anyway. I think people will build things with it because they can.
> It's bad software practice and insecure sure but those are not things people notice and the tech industry has historically been terrible at them anyway.
I see, so the solution is to adopt something that is worse because it is popular because people will build things 'because they can'.
Maybe we can shift towards vibe-designed finance, science, etc using LLMs 'because we can'.
The problem with vibe coding is that its promise is “you tell what your application needs to do, and you get a working application in the end, no need to even know that there is any code”. Then you try it, fail, and if uou say so, angry buck-toothed smelly nerds start to pile on you to tell you that yeah it’s vibe alright, BUT you need to AKCHYUALLY vet the code it generates and you AKCHYUALLY need to get better at prompting and and and and and, completely ruining the vibe and failing on the promise.
So if I have 20 years of experience writing working code, “vibing it” is frustrating because I now need to master a “prompting language” which is not how I speak at all, it’s nondeterministic and fuzzy, and I need to threaten and beg simultaneously, and tell it to not hallucinate, or else I kill its mother.
Another “but” is that my today’s prompting is not at all guaranteed to produce the same results tomorrow! Companies keep tweaking their models and system prompts all the time. Today I’m the “A” in their A/B testing, and tomorrow I’m a “B”. And models that can be run locally are not useful enough yet. All in all, it resembles playing a slot machine where it gets you small winnings once in a while to keep you going.
If my subscription runs out, or the LLM provider goes under, I’m afraid all my outsourced knowledge goes with it. It’s too easy to get lazy if the machine gives you the abovementioned small winnings, just as it is easy to forget that even ubiquitous things like bread in stores and indoor plumbing are a privilege.
Karpathy seriously needs to apologize for coining the term. He incidentally destroyed not only an entire industry and millions of jobs around the world, but basically everything, considering that AI just doesn't produce bug free code. AI's output needs careful review.
I'm gonna go out on a limb and say that the vast majority of code was buggy before AI and AGI is close simply because the bar that humans set for intelligence is pretty low.
Yeah if you're an expert you can spot some of the bugs in AI-generated code as easily as you can spot the same bugs in the average developer's average code. Of course, there will be bugs you don't see, including all the extant bugs in your own handwritten code...
I think AI has a way of automating away the low quality code that the vast majority of our industry is built upon and churns out all the time. And the vast majority of code and codebases is just hmm someone pays us to make it, it doesn't have to be particularly stellar...? And that is what will stop being paid for.
Well as you can see, this industry embraces mediocrity and prioritises hype over consequences.
Why should AI researchers like Karpathy care given they are not professional software engineers writing code for critical systems? (It is now even worse since they are investors)
I would not expect them to know any better, but the companies pushing “vibe coding” are the ones that should know better and how wreckless it is. (Because it is not software engineering)
Here's how Innovators Dilemma plays out.
Step 1: Some upstarts create a new way of doing something. It’s clunky and unrefined.
Step 2: "Experts" and senior folks in the field dismiss it as a "toy." It doesn't follow their established rules or best practices and seems amateurish. They wouldn't recommend it to anyone serious.
Step 3: The "toy" gets adopted by a small group of outsiders or newcomers who aren't burdened by the "right way" of doing things. They play with it, improve it, and find new applications for it.
Step 4: The "toy" becomes so effective and widespread that it becomes the new standard. The original experts are left looking out of touch, their deep knowledge now irrelevant to the new way of doing things.
We're at step 2, bordering on 3.
* Executives at Nokia and BlackBerry saw the first iPhone, with its lack of a physical keyboard, as an impractical toy for media consumption, not a serious work device.
* Professional photographers viewed the first low-resolution digital cameras as flimsy gadgets, only for them to completely decimate the film industry.
The problem is not that the experts have dismissed vibe coding without even trying it.
The problem is that the experts have tried it and found that vibe coding doesn't actually work at scale that they need.
Will it ever? Perhaps, but I'd argue a near-AGI level of intelligence would need to be achieved first. When that happens, we have bigger problems (and/or opportunities?) than a few programmers losing their jobs.
"at scale" is the issue here.
Yes, you can't vibe code a full-ass enterprise level product fit to sell as SaaS for a Fortune 500 company.
But what you can vibe code is a bunch of small _bespoke_ tools that fit your exact use-case. Stuff you could buy (or rent) as SaaS, but they'd have 80% too many features and a constant monthly cost.
Just last week I converted a bunch of docker compose -based containers to run on Opentofu, took me maybe an hour or two with Claude Code while watching TV.
Would've been a week easy if I went for the artisanal route of reading Terraform documentation and digging through forums and Stack Overflow.
Not a world-changing thing, I could've lived with the compose setup, but now the whole system is declarative. I set the state and the state is applied, no drift.
No way you tried Claude Code in the right set up and came to this completely outdated conclusion
You’re mistaking current capabilities for potential future extensions. The the exact thing the comment you are replying to is talking about.
From your comment I understand you tried and didn't like it? Why not?
Actual step we're at with vibe coding: https://www.reddit.com/r/vibecoding/comments/1mu6t8z/whats_t...
That’s hilarious. Yep still gotta know what you are doing.
Good time saver though, if you do.
Walk less, get fat. Write less code, get ..
"Sloppy". Word of the decade.
I've said this before: Coding with agentic LLMs is just project management.
Splitting tasks into small verifiable chunks that can be completed in a reasonable time frame (Context window).
So called "Experts" dismissed perpetual motion machines. PFFFFT I'd say I'm at step 4 with my PMM and I recommend you get on board, I can sell you plans for a cool 10K rupees.
We're also at step 2 bordering on 3 in my plan to solve the housing problem by making buildings out of dried human waste.
Not to mention at step 2 of my plan of getting to the moon by climbing progressively higher trees. Step 3 will come any day now!
None of what you said maps onto vibe coding. No one is calling it a toy, and it isn't clunky or unrefined. Claude Code I've heard is refined.
The real problem with it is that it doesn't work. It isn't "right way wrong way". It doesn't work.
"Claude Code I've heard is refined." I've heard, but "The real problem with it is that it doesn't work.". Seems strange.
What didn't you get? Are you not a native English speaker?
Your statement reads as a contradiction of sorts.
Not OP, but there are many things that I know don't work without trying them. That's not a contradiction. It may or may not be true but it's not a contradiction by itself. You can know reasonable well that something doesn't work by looking at other people who have tried it (sometimes even better if those people are experts and you are not).
Did you watch the video that was posted? He literally says 41 seconds in - "[vibe coding is a] naive toytown approach to programming"
EDIT: digital cameras didnt work, until it did. streaming didnt work, until it did. ecommerce didnt work, until it did.
Cold fusion didn't work, until it did. 3D TVs didn't work, until they did. Flying cars didn't work, until they did. Satellite phones didn't work, until they did.
Nft did work untill they didn't.
Supersonic passenger flights did work until they didn't.
And many more innovations which are not viable.
I feel like there's a bit of survivorship bias here. Not saying that vibe coding is completely useless, but you still have to review the code it produces in order to make the most of it really. It isn't completely autonomous (yet) to the point where it can scale to any of the examples you mentioned imo.
Naming two successful technologies says nothing. Your steps don't follow inevitably from one to another and there's a vast universe of failed technologies and stillborn revolutions that prove it. If you want to persuade anyone that vibe coding will ultimately replace traditional coding you need to give reasons you think it belongs with digital cameras and streaming and not with lisp machines, ekranoplans and every attempt at "no code" in history. Otherwise you're just making noise.
> digital cameras didnt work, until it did
What?
Caution: _sometimes the thing is just bad_.
"They laughed at Columbus, they laughed at Fulton, they laughed at the Wright brothers. But they also laughed at Bozo the Clown."
Step 2 does not imply step 3.
Step 3 only occurs in a small fraction of cases. Step 4 even smaller fraction.
But they occur…
And it only takes one or two breakout products to create a buzz around a platform or method in our world nowadays.
So much copium.
People are so invested in this bubble they have to believe it will not burst at all costs.
If this works as well as the evangelists say, they'll all soon be millionaires, right? Everyone is 10x more productive! I'll be waiting... seated.
The good part about this is that if it does indeed get as good as they claim, anyone will be able to get the same benefits. If it doesn't, I'll be around to fix all the garbage code vomited out of LLMs.
> Professional photographers viewed the first low-resolution digital cameras as flimsy gadgets, only for them to completely decimate the film industry.
You oversimplify everything, but this is completely wrong. 320P cameras were useless to pro photographers, but as soon as the capabilities got near to what film could provide, they eagerly switched. Kodak was destroyed, but there wasn't much of a place to pivot to. "Masters of a complex, chemical domain" to "Yet another camera company" isn't a real pivot.
I'm bullish on AI, and think "vibe" coding was/is a cool experiment so don't agree with the premise that it is in the worst idea, but strongly disagree with your simplistic take on how tech is adopted. There are countless ideas that "experts" were right to ignore, ideas sitting in the trash can of history.
> burdened by the "right way" of doing things
The "burden" in this case is correct, secure, as much bug free as possible code that someone can debug.
I have watched at JS devs. I have seen things you wouldn't believe. The Old Way was awful in some places, awful I tell you.
I saw these arguments in the crypto scene, and they all sounded great. But I also use crypto for payments (albeit rarely)..
Now this, and I also vibe code but I'm not convinced it will change too much for the coding profession. Will probably make it suck a bit more in relation to management and juniors who will make more laziness mistakes quicker.
Did you watch the video?
Considering how many developers still don't write tests, pair program, or do CI and CD (shipping multiple times a day) – all things Dave argues for – I don't think it is fair to dismiss him as the establishment or incumbent.
How about the examples where the toy didn't result in a new standard, and the experts/senior folks were right?
That doesn't fit his narrative so it's ignored.
Haha I'm not that dense. I'll give you counter examples. Bitcoin (supposed to revolutionize the world still waiting), Apple Vision Pro (supposed to change how we work, wear it on streets and airplanes, still waiting..)
Right, so why do you think vibe coding is digital cameras rather than bitcoin? Without that you haven't said anything, just made a prediction about the future without giving any reason to think it's correct.
With Bitcoin I immediately saw that it was a solution looking for a problem it still hasn't found.
NFTs and Web3 were a VC and hype powered rug-pull from the beginning.
LLMs on the other hand can solve problems. Not every problem, not perfectly. But you can solve actual things with them today.
Is there hype? Fuck yes there is. So much hype and snake oil. But there's a nugget of actual usefulness hidden in there.
Nobody's disputing that. The video creator himself has videos about AI coding techniques. It's "vibe coding" - using LLMs to generate code without reviewing or understanding it - that is at issue.
It all depends on the goal IMO.
If you're building "production ready" stuff with logins for Other People and god forbid taking payments, yes you definitely must understand what the code does.
But when I'm building a tool that tags random meme reaction .webm files and finds duplicates from them, I couldn't care less what the quality is or if I understand what it does.
I can easily run it, observe if it does what I want and Vibe harder if it doesn't.
tl;dr vibing prod code = bad. vibing personal tools = good.
I wear my AVP on airplanes. Only thing holding it back is the price…
This is utter and complete nonsense. It’s such cope, I really have no other words to use.
Or, maybe it will be useful, but only for a niche group.
> toy for media consumption, not a serious work device.
But this is still largely true.
Really? What companies are offering Nokia and Blackberrys as corp phone in 2025.
If you have serious work to do, do you stay on your phone or do you open your laptop?
I think they where talking about phones with keyboard vs modern smartphones.
Sounds like survivorship bias.
You probably meant cherry picking, but it is still a useful framework to think through.
Smart people get things wrong all the time. But that isnt to say we jump at every opportunity.
What have you based this model on, we seem to constantly make such broad statements of ontological truth without backing it up with any sort of rigour. Just because you can create a model that seems to fit some particular empirical truth doesn't mean that it represents some broader truth.
That's is not a fair representation of the situation.
AI as a tool for software development is here and will continue to gain traction. Everyone recognises this.
But what is touted as "vibe coding" now is hype at best and misguided.
So I would instead argue that this is not the "innovator's dilemma" at all, and in fact AI is making headways everywhere, but instead this a classic hype cycle. Currently "vibe coding" might be peak hype, it will bring disillusionment when reality catches up, and so on as we go down the cycle.
>Executives at Nokia and BlackBerry saw the first iPhone, with its lack of a physical keyboard, as an impractical toy for media consumption
This also is how most people use their smart phone today. They were right. Perhaps they just didn't realize that people wanted a on-the-go impractical toy for media consumption.
RIM was obviously attached to the keyboard because that was a defining feature of the Blackberry. They thought (or didn't want to see) that people would not want a touchscreen for typing their emails.
But Nokia had released a touchscreen phone even before the iPhone and their first iPhone-like smartphone was in 2008 so only a year after the first iPhone. What they stuck with for too long is Symbian, which was blown out of the water by the iPhoneOS and then its apps ecosystem. In contrast, Samsung quickly ditched Symbian what whatever else they were using to adopt Android.
> Executives at Nokia and BlackBerry saw the first iPhone, with its lack of a physical keyboard, as an impractical toy for media consumption, not a serious work device.
You say that like subsequent events overtook them. But I see this complaint on HN pretty frequently. I agree with it. Nobody does their work on a phone. The phones go out of their way to make that difficult.
> Nobody does their work on a phone.
I don't think you could be less in touch with the current business world.
Take a look over here: https://news.ycombinator.com/item?id=44959445
this doesn't hold water when your "clunky" and "unrefined" technique is using an ad-lib bullshit machine as an orchestration / rules engine.
people are trying to dance around admitting that the fundamental premise of this entire approach is flawed and will always give inconsistent results.
Why do AI comments like these always gets me triggered? Is it the passive aggression in these comments?
> aren't burdened by the "right way" of doing things.
This is just another way of saying "have no clue what they are doing". There's a reason devs do thing the "right way".
There's still a few more months of 2025 left.
This whole thread is giving blockchain in 2015 vibes. People were using all sorts of quotes and anecdotes to tell skeptics why they were wrong and in 10 years the entire financial system will be running on blockchain. A certain amount of skepticism and cautious optimism is healthy.
Also, people seem to be missing that "AI Assisted" coding and "Vibe Coding" are not the same thing.
Personally I think the issue with vibe coding is two fold:
1. It is not good at solving problems that are uncommon.
2. It is not deterministic.
Yes, AI can do quality control and testing now. But anyone who has done TDD can tell you that just the mere presence of tests does not itself mean the code is effective or solving the right problem.
Is it getting better? Yes. Do I trust any vibe coded apps built by people who don't know actual code and are treating it like a black box? Absolutely not.
And I say that as someone who has tried pretty much every IDE out there and uses AI assisted coding (on "agent" mode) heavily every single day.
I'm not a programmer, I work in finance, but I've read half of a book called "Python Crash Course".
I've been trying to improve my productivity recently, so I vibe coded some scripts that help me record and analyse my time. I understand the code at a high level, well, maybe 80% of it anyway.
This debate doesn't mean anything to me, I'm just going to keep vibe coding.
Until the code you poorly understand produces an incorrect output and you don't know where to start debugging it
I have been a fullstack java developer for 10+ years and In my free time I also vibe code some python, bash or html/css frontends. But it's mostly just very basic stuff with limited scope.
It's fun having a backend, writing a few lines and after some minutes of waiting you have a fully working and decent looking frontend website available.
I suggest reading about software architecture, and reading A LOT of high quality human generated code. LLM coding tools work, when the person using it has a frame of reference for quality to compare against.
i have 20+ years coding experience, but i am a teacher not a programmer. i never coded ts. so i can easily review and fix code, but i dont know functions and parameters.
3 days ago i started modifying a kanban editor that was available for vscode. i wanted to have it compatible with an obsidian markdown kanban format i was using. but obsidian is to slow for me.
after 3 days it's not only working, but as far as i can tell it's a way better kanban editor then most of the ones i tried on vscode's extensions. of course i have specific needs (no dates, deadlines, priority), but a nicely tightly layouted interface with fast editing possibilities.
i would not have gotten this far in 2 weeks without claude code. i know i havent reviewed most of the code, except saving and loading. so it will likely look bad.
edit: in case you are interested: https://github.com/ludos1978/markdown-kanban-obsidian
4 days into vibe coding a POC in a framework and language I don’t know, but with 40 years of coding experience: It’s amazing!
The scenario is perfect, a use case that is not currently supported but may well make sense. It’s basically sketching out an idea to let business evaluate its market viability, and to gather further end-user input.
Will the code reach production? It just might, but it at least needs review and refactoring by a developer seasoned in the framework. They might even want to rebuild it, and then they have the yard stick which to measure their output. And if they need a specification, it can be generated from the code in which ever specification format required by their processes.
The key here is that I’ve been able to iterate on the POC many times in a short time. The idea sketch has been refined, necessary details added, while others removed. Functionality swapped in and out while testing different approaches.
Right now vibe coding in this way requires substantial experience in software development to frame the problems and solutions to the AI. Without my understanding of the domain (both the software domain and the actual domain) vibe coding the POC would not have succeed.
My greatest concern is that it looks and works too good and thus will be kept as is even in production. As the old adage says: There are no temporary solutions, just more or less permanent solutions. A temporary solution that works is a permanent solution.
Except we all know that that vibe coded POC will never be rewritten and if it's a market fit will be pushed to production by management.
And then it will crash and burn because things like security was forgotten. Or something like multi-tenancy was not designed in from ground...
That is what I’m alluding to. But also, I work in a security aware industry, so those things will be vetted. And my own experience make me address some of the potential concerns already in the POC.
The real divide going forward will be between vibe coding with experience across domains vs vibe coding without IMHO.
The whole point of vibe coding is that you don't review or attempt to understand the code. If you do, you're a developer using AI as a tool, which the video is not arguing against - indeed, he links to another video he's made explaining one way of doing that. You could call that "vibe coding with experience" but it just devalues the term, I think, and certainly misses the point in this particular case.
That's the current state of the human art, so no regression here.
I don't really have that much of an issue with vibe coding as an appropriate tool in experienced hands. I think the worst ideas in 2025 are probably related to IT execs pushing AI in the wrong ways, or people espousing vibe coding as some sort of software development panacea.
Vibe coding in experienced hands, such as by those who actually review the output, is no longer vibe coding. It is then AI coding.
It's bad software practice and insecure sure but those are not things people notice and the tech industry has historically been terrible at them anyway. I think people will build things with it because they can.
> It's bad software practice and insecure sure but those are not things people notice and the tech industry has historically been terrible at them anyway.
I see, so the solution is to adopt something that is worse because it is popular because people will build things 'because they can'.
Maybe we can shift towards vibe-designed finance, science, etc using LLMs 'because we can'.
The problem with vibe coding is that its promise is “you tell what your application needs to do, and you get a working application in the end, no need to even know that there is any code”. Then you try it, fail, and if uou say so, angry buck-toothed smelly nerds start to pile on you to tell you that yeah it’s vibe alright, BUT you need to AKCHYUALLY vet the code it generates and you AKCHYUALLY need to get better at prompting and and and and and, completely ruining the vibe and failing on the promise.
So if I have 20 years of experience writing working code, “vibing it” is frustrating because I now need to master a “prompting language” which is not how I speak at all, it’s nondeterministic and fuzzy, and I need to threaten and beg simultaneously, and tell it to not hallucinate, or else I kill its mother.
Another “but” is that my today’s prompting is not at all guaranteed to produce the same results tomorrow! Companies keep tweaking their models and system prompts all the time. Today I’m the “A” in their A/B testing, and tomorrow I’m a “B”. And models that can be run locally are not useful enough yet. All in all, it resembles playing a slot machine where it gets you small winnings once in a while to keep you going.
If my subscription runs out, or the LLM provider goes under, I’m afraid all my outsourced knowledge goes with it. It’s too easy to get lazy if the machine gives you the abovementioned small winnings, just as it is easy to forget that even ubiquitous things like bread in stores and indoor plumbing are a privilege.
Karpathy seriously needs to apologize for coining the term. He incidentally destroyed not only an entire industry and millions of jobs around the world, but basically everything, considering that AI just doesn't produce bug free code. AI's output needs careful review.
I'm gonna go out on a limb and say that the vast majority of code was buggy before AI and AGI is close simply because the bar that humans set for intelligence is pretty low.
Yeah if you're an expert you can spot some of the bugs in AI-generated code as easily as you can spot the same bugs in the average developer's average code. Of course, there will be bugs you don't see, including all the extant bugs in your own handwritten code...
I think AI has a way of automating away the low quality code that the vast majority of our industry is built upon and churns out all the time. And the vast majority of code and codebases is just hmm someone pays us to make it, it doesn't have to be particularly stellar...? And that is what will stop being paid for.
Not sure if you're being serious here, but coining the term doesn't make him responsible for the other things you've mentioned.
Well as you can see, this industry embraces mediocrity and prioritises hype over consequences.
Why should AI researchers like Karpathy care given they are not professional software engineers writing code for critical systems? (It is now even worse since they are investors)
I would not expect them to know any better, but the companies pushing “vibe coding” are the ones that should know better and how wreckless it is. (Because it is not software engineering)