I am not a coder. I can't write any rule of python, javascript or c ++. With the exception of a short period in my teenage years when I built websites and tinkered with Flash animations, I have never been a software engineer, nor do I have ambitions to give up journalism for a career in technical industry.
And yet, in recent months, I have coded a storm.
Under my creations: a tool that transcribes and summarizing long podcasts, a tool to organize my bookmarks for social media in a searchable database, a website that tells me or fits a piece of furniture in the trunk of my car and an app called Lunchbox Buddy, which should get the contents of my fridge to decide.
These creations are all possible thanks to artificial intelligence and a new AI trend that is known as 'vibecoding'.
Vibecoding, a term that was populated by the AI researcher Andrej Karpathy, is a useful steno for the way in which today's AI tools even enable not -technical hobbyists to build fully functioning apps and websites, simply by typing instructions in a text box. You don't need to know how to cod to Vibode – just an idea and a little patience is usually enough.
“It's not really coding,” Mr. Karpathy wrote this month. “I just see things, say things, do things and copy stuff, and it usually works.”
My own vibecoding experiments are aimed at making what I call 'software for one' – small, tailor -made apps that solve specific problems in my life. These are not the kind of tools that a large technology company would build. There is no real market for them, their functions are limited and some of them only a kind of work.
But in this way buildings software-it describing a problem in a sentence or two, and then looking at a powerful AI model working on building an adapted tool to resolve it a stunning experience. It produces a feeling of AI Vertigo, similar to what I felt after using chatgpt for the first time. And it is the best way I have found to show the possibilities of today's AI models, which can now automate large chunks of basic computer programming and soon be able to use similar performance in other areas.
AI coding tools have been around for years. Earlier, such as Github Copilot, were designed to help professional codingers work faster, partly by finishing their code lines in the same way as Chatgpt completes a sentence. You still had to know how to cod to get the most out of it, and got in when the AI got stuck.
But in the past year or two new tools have been built to take advantage of more powerful AI models that can even program neophytes as professionals.
These tools, including cursor, replit, bolt and sweet, all work in similar ways. Given the prompt from a user, the tool comes with a design, definitely the best software packages and programming languages to use and works on building a product. Most products make limited free use possible, with paid levels that unlock better functions and the ability to build more things.
For a non-programmer, vibecoding can feel like sorcery. After you have followed your fast, mysterious code lines, flies past, and a few seconds later, if everything goes well, a working prototype comes to the fore. Users can propose tweaks and revisions, and when they are happy with it, they can implement their new product on the internet or perform them on their computers. The process can only take a few minutes, or as long as a few hours, depending on the complexity of the project.
This is what it looked like when I asked Bolt to build an app for me that could help me pack a school lunch for my son, based on a uploaded photo of the contents of my fridge:
The app first analyzed the task and broke it into components. Then it went to work. It generated a simple web interface, opted for an aid for image recognition to identify the food in my refrigerator and developed an algorithm to recommend meals based on those items.
If the AI needed me to make a decision – or I wanted the app to mention the food fees of the food that it had recommended, for example – I asked me with different options. Then it would go out and cod a little more. When it got a snag, it tried to debug his own code, or went back to the step before it got stuck and tried another method.
About 10 minutes after I had been introduced my promptly, Lunchbox Buddy – that was what the AI had decided to call my app – was ready. It suggested a generic sandwich with turkey. You can try it yourself here. (The version I have built contains an AI image recognition tool that costs money to use; For this public web version I have replaced it with a simulated image recognition function, so I don't get a huge bill.)
Not all my vibecoding experiments have been successful. I have struggled for weeks to build a “Inbox Autopilot” tool that can automatically respond to my e -mails in my writing style, in my writing style. I encountered roadblocks in integrating AI work flows in apps such as Google photos and iOS speech memos, which are not designed to play well with add-ons from third parties.
And of course AI occasionally makes mistakes. Once, when I tried to build a website for a tire shop in my neighborhood, invented the AI fake reviews of the Yelp page of the store and added to a testimonial page. Another time, when I tried to turn a long story, I had written an interactive website, the AI contained about half the text and omitted the other half.
In other words, vibo coding still benefits from having people who supervise the robots, or at least float in the neighborhood. And it is probably the best for hobby projects, no essential tasks.
That may not be much longer. Many AI companies are working on software engineering agents who can fully replace human programmers. AI already reaches world-class scores on competing programming tests, and various large technology companies, including Google, have outsourced a large part of their engineering work to AI systems. (Sundar Pichai, the Chief Executive of Google, recently said that AI-generated code was more than a fourth of all new code from Google.)
If I was a junior programmer – the kind that AI seems to be most likely to replace – I might be able to panic about my work perspectives. But I am just a man who likes to tinker and build up aids that improve my life in small ways. And vibo coding – or actual coding – is an area where AI unmistakably improves.
Since I talked about my vib code experience on my podcast last week, I have heard of dozens of other people who have built their own tools with AI -Help. Colleagues told me about the food apps they have built to help them stick to their diet, or the tools they use to summarize the E -mail newsletters they get. Readers have sent websites that they have built to keep track of the price of eggs, or scraping Zillow listings in Los Angeles to discover rental yielding authorities after the fire of Palisades.
Little of these tools are in themselves world changing. What is new and remarkable is that amateurs with a few test attacks can now build products that would have previously required teams of engineers.
I am not Pollyannaish about AI, or blind to the effects that AI coding apps could have on society if they continue to improve. I think it is possible that an AI that automates the construction of useful software can also automate the creation of malignant code, or can even lead to autonomous cyber attacks. And I am worried that software engineering is only the first known profession that the labor-falling effects of AI tools experience.
But for the time being, building apps seems to automate annoying or time -consuming tasks in my life, a just as good use of AI as one. So I keep vibecing – at least until my child can pack his own lunch.
