Will 'Vibe Coding' Transform Programming? (npr.org)
- Reference: 0177887991
- News link: https://developers.slashdot.org/story/25/06/01/0251235/will-vibe-coding-transform-programming
- Source link: https://www.npr.org/2025/05/30/nx-s1-5413387/vibe-coding-ai-software-development
NPR [1]explores vibe coding with Tom Blomfield , a Y Combinator group partner:
> "It really caught on, this idea that people are no longer checking line by line the code that AI is producing, but just kind of telling it what to do and accepting the responses in a very trusting way," Blomfield said. And so Blomfield, who knows how to code, also tried his hand at vibe coding — both to rejig his blog and to create from scratch a website called [2]Recipe Ninja . It has a library of recipes, and cooks can talk to it, asking the AI-driven site to concoct new recipes for them. "It's probably like 30,000 lines of code. That would have taken me, I don't know, maybe a year to build," he said. "It wasn't overnight, but I probably spent 100 hours on that."
>
> Blomfield said he expects AI coding to radically change the software industry. "Instead of having coding assistance, we're going to have actual AI coders and then an AI project manager, an AI designer and, over time, an AI manager of all of this. And we're going to have swarms of these things," he said. Where people fit into this, he said, "is the question we're all grappling with." In 2021, Blomfield said in a podcast that would-be start-up founders should, first and foremost, learn to code. Today, he's not sure he'd give that advice because he thinks coders and software engineers could eventually be out of a job. "Coders feel like they are tending, kind of, organic gardens by hand," he said. "But we are producing these superhuman agents that are going to be as good as the best coders in the world, like very, very soon."
The article includes an alternate opinion from Adam Resnick, a research manager at tech consultancy IDC. "The vast majority of developers are using AI tools in some way. And what we also see is that a reasonably high percentage of the code output from those tools needs further curation by people, by experienced people."
NPR ends their article by noting that this further curation is "a job that AI can't do, he said. At least not yet."
[1] https://www.npr.org/2025/05/30/nx-s1-5413387/vibe-coding-ai-software-development
[2] https://www.recipeninja.ai/
I read the article (Score:5, Interesting)
What do you get when you take some naive kids pursuing the American dream, add Y Combinator Kool Aid, and an NPR reporter who needs to fill a page with hopeful words?
30,000 lines of code ? For a recipe website? What effing language(s)? Bloated much? 500 lines of CRUD and 29,500 lines of garbage?
Re: (Score:1)
> What do you get when you take some naive kids pursuing the American dream, add Y Combinator Kool Aid,
$500K buys a lot of Kool-Aid.
> 30,000 lines of code ? For a recipe website? What effing language(s)? Bloated much? 500 lines of CRUD and 29,500 lines of garbage?
If you measure Bloat in lines of code, yes. But if you measure Bloat in the amount of time and effort needed to maintain and extend, maybe not. Not that AI has achieved that goal yet, but it wouldn't hurt for us to all start using the right metric, so we'll know a good thing when we see it.
Re: (Score:2)
If you're saying this is a self perpetuating cycle of give me money and don't audit the outputs, then yes, I agree.
Also, please note as I pointed out below, that this website was pre-built in several hundred WordPress templates.
So, this whole story and website is pure make work for the sake of ... you know.. "the economy".
Re: I read the article (Score:2)
Oh what is CRUD? Being capitalized I assume it is an acronym.
Re: (Score:2)
Create read update delete
Yes, but no.. (Score:4, Insightful)
..At least in near future.
AI will very likely/definitely replace many "coders". That is people that got a mission of "Code a function that does X" and then would copy-paste code from online sources and slightly modify them.
There are supricingly high numbers of those people. They do not have to understand things, just implement a functionality that is clearly defined.
Another group that is in danger now/soon are people who do simple reports and similar programming that is basically short programs that do a single functionality.
We are still pretty far from replacing people who actually know what to do in more complex scenarios, though their work will most likely be assisted quite a lot by the ability to quickly get the types of things listed above done instead of typical handing off to others in the organization.
In the longer run, no idea when and to what extent the systems will be able to tackle harder problems, that remins to be seen.
Re: (Score:3)
It's not clear to me that the people making the hiring/firing decisions, and deciding how many programmers can be replaced by AI, know the difference between the copy-paste coders you're talking about and the people who are doing the harder things.
And, given the way they think, and given the fact that all of us are subject to a whole host of cognitive biases, some places at least are likely to want to keep on the cheap copy-paste types than the more expensive senior programmers.
Short term, things will look
Re: (Score:2)
I'm not sure even if it will cost less. I saw some vibe coding sessions and I'm sure it was only being done because the user had the top end account and likely the whole session (which ended unsuccessfully) probably cost several thousand dollars in compute time.
Simple requests that everyone do, sure, can probably be done for less money. But once you need something revised it will probably take many iterations and start racking up the bills.
And before you know it, after a week, you've racked up tens of thous
Yes. 100% (Score:5, Insightful)
It will definitely transform programming and not in a good way.
Transform? Yes. (Score:2)
But like the article says, you still need humans to comprehend and maintain the code. AI can generate a lot of code in a short amount of time - it may not be the most efficient, or even viable, and that is where expert humans come in.
Re: (Score:2)
Maintain code?
The LLM can just regenerate it from scratch almost for free.
Re: Transform? Yes. (Score:2)
So instead of maintaining, just roll the dice again and hope the AI is perfect next time?
Re: (Score:2)
That reminds me of my O(N!) sort algorithm. (Really, it was a student of mine who proposed this, as a joke.)
(1) Randomize the array
(2) Is it sorted? If not, goto (1).
Re: (Score:2)
It doesn't have to be perfect, it only has to be as good as an average programmer.
Not transforming programming but entrepreneurship (Score:4, Insightful)
It enables to quickly come up with a working prototype for an app idea. If you get funding you hire engineers to rewrite it the proper way.
Re: (Score:2)
For reasons I won't go into beyond saying it's an educational projects for kids, I needed to write a (very simple) app for an iPhone 3G recently, Problem is: I've never written any kind of app at all. I have remnants of a single college level C++ class from thirty years ago, that's it.
I was lucky to have a friend 'in the business' - he pointed me in the right way and got me set up in Xcode (4.4, in case you're skeptical), but I was astonished how much AI (specially Chat-GPT) was able to pull me through. Exp
Please let's not force them (Score:2)
to use a top down management style! What are you trying to do, create AI Dilbert?
People have gotten funding without a line of code (Score:2)
It changes programming for sure. But I don't really get this story much.
> A 21-year-old's startup got a $500,000 investment from Y Combinator — after building their web site and prototype mostly with "vibe coding".
Doesn't really mean much. Visual programming, low code & no code is not new and have all been funded. Getting funding on just an idea and a team (but no code) happens all the time. Y-Combinator acceptance these days is all about being a rich kid with Ivy League background, and having a som
Re: People have gotten funding without a line of c (Score:2)
WordPress has a hundred templates already built and ready to go, right now. No programming or vibes needed. You're welcome. Give me $500,000 :-)
Betteridges Law (Score:4, Interesting)
No.
But some people will use it to their benefit. Did you try GLM? It write you a paint application in one prompt. Without programming skills you won't get it to become photoshop, but if you wanted a free and ad free paint for your iPad (with some smaller features that you were missing from other paints) its great. If you aren't a programmer you will easily get stuck. Still many of the one-shot apps will help you. You want a tool that can batch-sharpen your images? No problem, even as non-programmer if you can install python and prompt a model you have your solution. If you think you will get a programming job with that you're misguided.
AI raises ceiling and floor. Professionals were not on the floor, but the good ones can now be at a higher ceiling. Who previously had to rely on others to help out with scripts to batch-process images, can now do it themselves. As always, some of the solutions will be suboptimal, cumbersome and professionals will shake their heads how they are less straightforward than they could be. But they work for the users and that's what matters.
If a user learns to achieve a goal with 100 clicks it causes me headaches how one can do this without getting insane, still it is a working solution for them they use everyday because it achieves their goal better than the existing tools and they are unable to program the better solution. Now they will use less than optimal prompted programs, but they still won't need 100 inefficient clicks, but just run one inefficient script that would cause me headache, but is faster than the strange ritual they used before.
I wonder (Score:2)
how awful the recipes concocted by the AI are. I doubt they're as good as the AI-generated code, as bad as that may be.
AI Coding (Score:3)
My experience, mainly with generating SQL queries, is that AI inevitably gets it wrong multiple times, so what I have had to do is more of a kind of meta-programming; giving the model cues and corrections. I have created some pretty sophisticated SQL queries, but there's no way in hell I can just pop the first go-around into my code and have it run. Either it's outright faulty code that will fail, or it's just not producing the correct results.
Now SQL is a fairly limited and ring fenced language (excluding stored procedures of course). I've never tried it with a general use language, but I imagine those problems will get more pronounced. That's not to say it might not be useful for translating natural language specs into code, but if my experience with SQL is any indicator, it's going to require a lot of massaging. There's probably still productivity boosts to be found here, which will likely have in effect on the number of programmers out there, but to me, it feels more like a layer of abstraction that will require a different kind of programming, rather than replace programming.
As an example that isn't coding, I have been building models for what I expect is a government procurement next year. This involves taking previous Requests For Qualifications documents, updating them with current knowledge of government expectations, procurement rules, and so forth. Again, building these model RFQs is an iterative process, not simply one of "Take these RFQs from previous procurements, update them with this new information I've uploaded, and give me model RFQs based on these premises I will provide." My test run took about three or four hours of a kind of conversation, where I correct and shape, understanding the cues the LLM needs to produce the desired result, and the better I get at understanding not just the kind of information and cues the LLM requires, but the most effective means of "encoding" that information, the more efficient the LLM is at producing the desired results.
That sure sounds like programming to me, albeit at a much higher level of abstraction. LLM, at least where it stands, is just another platform, a very powerful one, but as with all programming languages, the larger the command set and the more complex the lexical structures, the more room for bugs, and the more subtle some of those bugs can be.
Re: AI Coding (Score:2)
Anecdotal, but one of our younger team members who's still at uni told us that the only lecture without rules against LLM use for programming exercises is the database one, because all the models fail to generate useful SQL anyway.
That being said, I sometimes try it for TypeScript and C#, and don't usually get results I'd consider useful either.
Re: (Score:2)
I have definitely produced useful SQL code, and indeed some pretty darned complex queries for transformations and data hygiene, but as I said, it's not a process of "dump spec into LLM model, run SQL on RDBMS", but rather a kind of meta-programming conversation. I imagine specialized LLMs might do a bit better, but I'm generally pretty skeptical of the current generations of AI building sophisticated software. I suspect where LLM's might do well is with interop code, the kind of boiler plate code that takes
Re: (Score:2)
I have the opposite experience, but I've probably been building and running databases for longer than your coworker has been alive.
Not a new thing really. (Score:4, Insightful)
We have seen this before with Visual Basic and PHP. Novices quickly churning out vast amounts of almost working code that can not be repaired, only replaced. This is just the next generation of the same problem. If you like re-implementing someone elses buggy software in a proper way, you got a bright future ahead.
Prototypes are the easy part (Score:2)
At my current company, I whipped up a prototype SSO application in three days. Then I managed a team that turned it into production-ready code in just...6 months. And that was a very small, special-purpose application that just does the basics of authentication, such as login, reset password, change password, and so on.
Yes, AI is great at whipping up prototypes. But if my own interaction with AI coding is any indication, AI struggles to harden code or make it robust. For that matter, half the time the code
Re: Prototypes are the easy part (Score:2)
I imagine coding for a single stage to orbit application must be very tricky. Why do people casually pepper their comments with their personal lingo incomprehensible to anyone outside their specialty?
Re: Prototypes are the easy part (Score:2)
Uh huh, never heard that in my life and I bet that's the case for the majority of people.
Re: (Score:2)
SSO = Single Sign On. It even has a Wikipedia page.
[1]https://en.wikipedia.org/wiki/... [wikipedia.org]
[1] https://en.wikipedia.org/wiki/Single_sign-on
The website has major problems (Score:4, Insightful)
So, I decided to try out this Recipe Ninja site that was so quickly drummed up with AI.
First reaction, it looks nice enough.
So, I clicked the microphone and said "fettuccini recipes." It correctly heard what I said and showed me a Search Results page with "Active Filters" listing "fettuccini recipes." In the middle of the page, it said "No recipes found matching your search criteria."
About a minute later, a voice came from the page saying, "I found some recipes for you, would you like to see the first one?"
I said, "Sure."
Nothing came after that, either visually or in audio form.
So in 2 minutes, I personally found some serious issues with the site. Is AI going to fix those issues? I doubt it.
It's easy to create a prototype. Even AI can do that. Can it harden the site and fix its bugs? I doubt it.
On a practical level, (Score:2)
The recipe for gnocchi puttanesca in a bag has an AI illustration of a meal cooked in a burlap drawstring bag. The bag in the actual Jamie Oliver recipe is a typical pouch made of foil.
Yes and no (Score:2)
Certainly a useful tool. In knowledgeable hands. Currently, you are still likely to need an experienced human to tweak and curate the result. Certainly to maintain it.
The wild card is how good the LLMs are going to get. We just don't know. They may hit a wall and stay there. Or they may not.
No (Score:3)
It will produce unmaintainable, insecure and unreliable crap for a few years and then it will quietly die.
Re: (Score:2)
I'd be careful of this attitude. It strikes of Thomas Watson's, "Only five computers worldwide," and the obsolescence of silent movies, among other examples of trends that were considered passing fads that actually became societal revolutions.
Product Liability (Score:2)
At some point something bad will happen and the company will be sued and they will claim they didn't even write the code and the AI vendor will say it's not their fault either.
Or the bridge their AI designed.
"Oopsie! "
I wouldn't trust the Courts to get any of this right either.
We'll maybe see contracts and insurance enforce human oversight but then people will cheat and settle for less than profits.
As they say, profit has replaced survival as the Human evolutionary fitness function.
It will certainly transform zero days (Score:2)
Its is really impossible for me to believe this concept will result in zero exploitable bugs. More likely it will infest every part of the process. What we need though is a nice cool name to go with 'vibe coding'?
Jive Coding (Score:2)
There, fixed that for ya.
It already did (Score:1)
Programmers should get into vibe-coding as much as possible, or they will be replaced quickly by other programmers who do.
Knowing how code works is a great benefit, and vibe-coders that have no real programming knowledge won't get over basic tasks. Programmers, with in-depth knowledge, definitely are more productive using vibe-coding for huge parts of a program, then programmers not using vibe-coding. The fine details, very carefully crafted specific things, can also be explained while vibe-coding. This can
Boeing Is Already Using It (Score:2)
As part of the effort to improve the chain of development,
Boeing is already replacing the outsourced engineers with "vibe coding,"
They started to implement the next iteration of the 737 MAX with and improvement in the MCAS system completely rewritten with this new magic tool.
Real point of recipe tale? (Score:2)
You can already just ask your favorite AI "get me a recipe for X" or "give me a recipe for these ingredients". Next, whatever that kids idea was, again, user could just go to one of the AIs and ask it directly. Don't need the kid, don't need the website.
Cmon people.
For simple stuff, maybe (Score:2)
Complex, novel software is complex, irreducibly complex. It can't be completely specified in a simple text prompt.
The prompt "write a snake game in python" works because snake games exist along with their open source code.
Managers want to reduce costs and dream of not needing to hire programmers. This will result in a tsunami of crappy code generated cheaply.
The real promise of AI tools is as assistants that will help experts analyze and manage complexity
What about 'new' stuff (Score:5, Insightful)
AI might replace existing programmers because it has been trained on what they have already done before, but what happens when something new comes along? A programmer faced with a new language, OS, API or whatever has to sit down and learn it from documents, not existing examples. Without programmers programmers creating stuff with the new thing there is nothing for the AIs to be trained on.
AI may be bringing changes, but it's limitations will become pain points for some people in the future.
Re: What about 'new' stuff (Score:2, Insightful)
I thought rust was all we needed from now on.