I’m not coding. I can’t write a single line of python, javascript or C ++. With the exception of a short period in my teens when I built websites and entered with flash animations, I was never a software engineer, nor have I have the ambition to give up a journalist for a career in the technology industry.
And yet, for the last few months, I have been encrypted a storm.
Among my creations: a tool that transcribes and summarizes long podcasts, a tool to organize my social media bookkeepers on a searchable database, an online site telling me if a piece of furniture will fit into my car luggage and an app called Buddy Buddy, which analyzes my refrigerator contents Tim.
These creations are all thanks to artificial intelligence and a new trend of the one known as “vibecoding”.
Vibecoding, a term that was popularly popular by Ai scholar Andrej Karpathy, is useful for how today’s tools to allow even -technicians to build fully functioning applications and websites, only by pressing requests in a text box. You don’t have to know how to encode in vibecode – just have an idea and a little patience, it’s usually enough.
“It’s not really coding,” wrote Mr. Karpathy this month. “I just see things. I say stuff. I run stuff and copy paste items, and it works mainly.”
My Vibecoding experiments have intended to do what I call “software for a” – small, hand -made applications that solve specific problems in my life. These are not the types of tools that a large technology company would build. There is no real market for them, their features are limited and some of them just kind of work.
But building the software in this way-the description of a problem in a sentence or two, then looking at a powerful model it goes to work by building a custom tool to solve it-is an experience in mind. It produces a feeling of that vertigo, similar to what I felt after using chatgt for the first time. And it is the best way I have found to demonstrate skeptics of today’s models of models, which can now automate large pieces of basic computer programming, and may soon be capable of similar deeds in other areas.
He has existed for years. Previously, like Github Copilot, they were designed to help professional coders work faster, in part finishing their code lines in the same way that the chatgt complements a sentence. You still have to know how to encode as much as possible from them, and open when he got stuck.
But over the past year or two, new tools have been built to take advantage of the most powerful models of one that enables neophytes to also program as pro.
These tools, which include the cursor, reproduce, lightning in the sky and loving, all work in similar ways. Given a user quickly, the tool comes up with a design, puts on the best software packages and programming languages to use, and gets to work on building a product. Most products allow free limited use, with paid levels that unlock better features and the ability to build more things.
For a non-programming, vibecoding can feel like magic. After you write your quick, mysterious code lines fly, and a few seconds later, if everything goes well, a work prototype appears. Users can suggest tweaking and reviews, and when they are satisfied with it, they can place their new product online, or direct them to their computers. The process can only take a few minutes, or as long as a few hours, depending on the complexity of the project.
Here’s what it looked like when I asked Bolt to build an app that could help me pack a school lunch for my son, based on a loaded photo of my refrigerator content:
The app for the first time analyzed the task and destroyed it in the constituent parts. Then got to work. She generated an underlying interface online, chose an image recognition tool to identify foods in my refrigerator, and developed an algorithm to recommend foods based on those items.
If he needed me to make a decision – if I wanted the app to list the nutritional facts of the foods he recommended, for example – this prompted me with several options. Then he would leave and cod even more. When she hit an obstacle, she tried to debug her code, or rested on the step before she had stuck and tried another method.
About 10 minutes after I had entered my quick time, Buddybox Buddy – that is what he had decided to call my app – was ready. She suggested a general turkey sandwich. You can try it yourself here. (The version I built includes an image recognition tool that costs money to use; for this public version online, I have replaced it with a simulated image recognition feature, so I don’t make a big bill.
Not all my Vibecoding experiments have been successful. I have been fighting for weeks to build a “Inbox Autopilot” tool capable of responding to my electronic posts automatically, in my writing style. I have encountered road blocks when I try to integrate the workflow of it into apps like Google Photos and iOS Voice Memorandums, which are not designed to play well with third -party additions.
And, of course, he occasionally makes mistakes. Once, when I tried to build a website for a tire store in my neighborhood, it constituted fake ratings from the Yelp page of the store and added them to a testimony page. Another time, when I tried to turn a long story I had written on an interactive website, it included about half the text and left the other half.
Vibery, in other words, still benefits from people who supervise robots, or at least hovering nearby. And it is probably the best for hobby projects, not the essential task.
This may not be true for much longer. Many companies are working on software engineering agents that can completely replace human programmers. Now, he is reaching world -class results in competitive programming tests, and some major technology companies, including Google, have transferred a large part of their engineering work to the Systems. (Sundar Pichai, Google’s chief executive, recently said the code generated by him constituted more than a quarter of the entire new code set on Google.)
If I were a new programmer – the type he seems to be most likely to replace – I can be in panic about my work prospects. But I’m just a guy who likes to show, and build tools that improve my life in small ways. And vibecoding – or actual coding – is an area where it is undoubtedly improving.
Ever since I talked about my vibecoding experience in my podcast last week, I have heard from dozens of other people who have built their tools with help. Colleagues have told me about the food applications they have built to help them adhere to their diets, or the tools they use to summarize the emails they receive. Readers have sent to websites they have built to follow the price of eggs, or to destroy the Zillow lists in Los Angeles to detect renting cases after Palisades fire.
Few of these tools are changing in the world in their right. What is new and noticeable is that with some keys, amateurs can now build products that would have previously searched for engineers.
I am not pollyannaish about him, or blind to the effects that he could have in society if they continue to improve. I think it is possible that one that automates the construction of useful software can also automate the creation of a malicious code, or even lead to autonomous internet attacks. And I worry that software engineering is just the first white collar profession to experience the effects of replacing it.
But for now, building apps to automate annoying or taking time in my life looks just as good as any. So I will continue to vibecoding – at least until my baby packs his lunch.