I'm not a coder. You cannot write a single line in Python, JavaScript, or C++. I was not a software engineer except for the short time I was a teenager who created websites and tinkered with flash animations.
Still, for the past few months I've been coding storms.
Some of my works are tools to transcription and summarise long podcasts, tools to organize social media bookmarks into a searchable database, websites that show whether furniture fits the trunk of my car, and help me to analyze the contents of my fridge and decide what to pack for my son's lunch.
All of these works are possible thanks to artificial intelligence and the new AI trend known as “Bibecord.”
VibeCoding, a term popularized by AI researcher Andrej Karpathy, is a shorthand for allowing even non-technical enthusiasts to build fully functional apps and websites, even today's AI tools, by simply typing a prompt into a text box. You don't need to know how to code it in a bibecord. Just having an idea is enough to have a little patience.
“It's not really coding,” Karpathy wrote this month. “I just look at things, say things, run your luggage, copy pasting, and most of it works.”
My own Bibecord experiment is aimed at creating what I call “Software for One.” These are not the types of tools that large companies build. They have no real markets. Their functions are limited, and some of them are their only job.
But building software like this – explaining the problem in one or two sentences, then watching a powerful AI model will lead you to build a custom tool to solve it – is an amazing experience. It creates the feeling of Ai Vertigo, just like what I felt after using ChatGpt for the first time. And that's the best way I've found to demonstrate to skeptics the capabilities of today's AI models. This will allow you to automate large chunks of basic computer programming, and may soon be able to do similar feats in other fields.
AI coding tools have been around for years. Previous ones like Github Copilot were designed to help professional coders work faster by completing lines of code in the same way ChatGPT completes statements. You still had to know how to code to get the most out of them and intervene when the AI got stuck.
However, over the past year or two, new tools have been built, leveraging more powerful AI models that allow even Nophytes to program like PRO.
These tools include cursors, replicas, bolts and lovely things, all work in a similar way. Taking into account user prompts, this tool will create designs, determine the best software packages and programming languages to use, and build the product. Most products allow limited free use with paid tiers that unleash better features and the ability to build more.
For non-programmers, bibecode can feel like magic. After entering the prompt, the mystical codeline line passes by, and after a few seconds, if everything goes well, a working prototype appears. Users can suggest tweaks and revisions. Also, if you are satisfied, you can deploy your new product to the web or run it on your computer. The process can take several minutes or hours depending on the complexity of the project.
What does this look like when I asked Bolt to an app that would help me stuff my school lunch for my son based on uploaded photos of the contents of the fridge?
The app first analyzed the tasks and split them into component parts. After that, it had to work. We developed an algorithm that generates a basic web interface, selects image recognition tools that identify foods in the refrigerator, and recommends diet based on those items.
If AI needs to make a decision on me – I was encouraged to have a few options, such as whether I would like to list the nutritional facts of food that I had recommended on the app. Then it turns off and code a little more. When it was caught, it tried to debug its own code or back up to a step before it got stuck and tried another way.
About 10 minutes after typing the prompt, the lunchbox buddy (the one that AI decided to call my app) was ready. It suggested a common turkey sandwich. You can try it yourself here. (The version I built has built-in AI image recognition tools that are expensive to use. This public web version replaces it with a simulated image recognition feature, so it doesn't create a huge invoice.)
Not all of my Bibecord experiments have been successful. I've been struggling for weeks to build a “Inbox Autopilot” tool that can automatically respond to emails in my writing style. When I tried to integrate my AI workflow into apps like Google Photos or IOS Voice Memos, I encountered a failure that was not designed to work well with third-party add-ons.
And of course, AI sometimes makes mistakes. Once upon a time when I was trying to build a tire shop website in my neighborhood, AI created fake reviews from the shop's Yelp page and added them to the testimony page. At another time, when I tried to talk about a long story I wrote on an interactive website, AI ruled out the other half, including about half of the text.
In other words, bibecords still benefit from humans overseeing the robot, or at least hovering nearby. And it is probably perfect for hobby projects and not an essential task.
It may not be true for much longer. Many AI companies are working on software engineering agents that can completely replace human programmers. Already, AI has achieved world-class scores in competitive programming tests, and several major tech companies, including Google, outsource the majority of their engineering work to AI systems. (Google CEO Sundar Pichai recently said that more than a quarter of all the new code deployed on Google is made up of.)
If I was a junior programmer – it seems most likely that kind AI will be replaced – I might panic about my job outlook. But I'm someone who just likes to tinker and build tools that improve my life in small ways. Bibecord, or actual coding, is one area where AI is definitely improving.
Since talking about my Bi-Becoded experience on my podcast last week, I've heard from dozens of other people building their own tools with AI support. My colleagues spoke to me about the nutrition apps they built to help them stick to their diet, or the tools they use to summarise the email newsletter they get. Readers have discovered cases of post-fire rent adjustments in Palisades by sending websites built to track egg prices and scraping off Zillow lists in Los Angeles.
Most of these tools are changing worldwide in and of themselves. New and noteworthy, with some keystrokes, you can now allow amateurs to build products that previously needed a team of engineers.
I'm not about polyannan about AI, and I'm not blind to the impact that AI coding apps have on society as they continue to improve. I think AI that automates constructive software can automate the creation of malicious code and lead to autonomous cyberattacks. I am also concerned that software engineering is just the first white-collar occupation to experience the labour-expressing effects of AI tools.
But for now, building an app to automate annoying and time-consuming tasks in my life seems to be using AI the same way as anything. So I'll continue with the Bibecord – at least until my child can pack his own lunch.
