Secret CEO In 1991, when I was 16, a Norwegian exchange student gave an inspiring performance of the Three Billy Goats Gruff, in the original Norwegian language, at my high school talent night. She delivered this performance with such enthusiasm that every word of her performance stuck in my mind and to this day I can recite the Three Billy Goats Gruff in Norwegian.
I can “Vibe Code” Norwegian.
I don’t speak the language, but that didn’t stop me from confidently using this skill with any Norwegian person I met. My parlor trick immediately falls apart as soon as they respond to me in anything other than English, but over the years I’ve used it as an icebreaker with the reserved people of Norway, because they find my heavily Australian-accented rendition of their culturally significant fairy tale cute.
This is the same reaction I had when I showed my freshly built package to our CTO, proudly declaring that I had decided to run the functional specification and user story of this new file project we were working on through an AI coding agent. The idea was to see if this would be useful to the project.
He asked me a series of pointed questions which immediately reminded me of the emotion I had felt when the poor Norwegian whom I had just delighted with my talent had responded with “Snakker du litt norsk?” (Do you speak a little Norwegian?) after which I was immediately perplexed and a little embarrassed. By using the word “litt” in the sentence, they were informing me that they knew I understood very little of what I was saying, but that they appreciated the effort.
Back to my conversation with the CTO, who looked at my ambiance-coded project and asked “Why is linting disabled here?” »
I wasn’t sure so I replied, “What does linting mean?” The CTO told me to hand over my laptop and go face the wall in the hardcoded credentials corner. “But I need it, I’m helping,” I protested.
“You will get it back when you realize what you did and apologize,” the CTO responded.
This wasn’t my first foray into Vibe Coding; I have been responsible for large-scale custom software projects for 20 years. I’ve been using story-driven software design for 12 years and have experienced several waves of software specification processes, from traditional functional specification, through behavior-driven design and testing.
I even had a short fling with Gherkin, mistakenly thinking that it would act as a happy medium between how developers and business owners would think about how to describe required functionality in software.
I felt like I was better equipped than most to tackle storytelling-driven development using AI. I had prepared skills, a long and varied catalog of reference projects, all using a strictly enforced entity library, security models and a common approach to schema definition.
I also had a string of successes under my belt where I used the AI coding tool to build some pretty impressive prototypes that received the appropriate number of oohs and aahs in some meetings full of people I was trying to impress. These prototypes turned into real projects and, intoxicated by a newfound confidence in the tools I was using, I turned my attention to transforming one of the fundamental concepts of the prototype into a real component.
It worked…until it didn’t.
Sober up
Here’s the lesson: the rhetoric around AI coding agents signifying the end of software development as a career is greatly exaggerated.
I have no doubt that one day humans will no longer type code line by line, but who has done that in recent years anyway? Stack Overflow really missed a trick by not charging every time a user used Ctrl-C on the site. This would have resulted in torrents of money, as millions of developers around the world realized that it is quite rare that a question asked has not been answered by someone else before.
Software Engineer Reveals AI Coding Assistants’ Dirty Little Secret: They Don’t Save Much Time
LEARN MORE
Copy-and-paste development shared many of the problems we see in Vibe Coding, because those who couldn’t understand the code they were about to use CTRL-V in a project should never have used it in the first place.
At least in the world of Vibe Coding, when you ask the AI to explain why it does something in a particular way, it doesn’t insult you, imply that it knows your mother better than seems possible, and explain to you why your n00b question is unworthy to answer.
Vibe Coding is a valuable skill to have. The value is amplified when you know what limitations to apply to your project. Experienced software developers immediately understand what these limitations and edge cases are.
The happier it makes you while you’re using it, the more you use it. Just like social media, it doesn’t matter if it’s true, it just matters that you stay face down in the feeder.
In experienced hands, vibe coding speeds up the development process so significantly that it certainly has a disruptive effect. Is it disruptive in the sense of spelling the end for developers? No way.
This is a well-defined economic paradigm, in fact chapter 7 of Joseph Schumpeter’s 1942 book Capitalism, socialism and democracy introduces the concept of creative destruction, which gives us a blueprint for how this will play out.
In the United States in 1970, the thriving telecommunications industry employed 420,000 switchboard operators, most of them young women, who manually connected 9.8 billion long-distance calls per year, an average of 64 calls per day per operator.
The invention of the auto attendant had a catastrophic effect on operator personnel. However, this also had a corresponding effect on the number of calls made (106 billion in 2000), prompting businesses to look for ways to manage the number of phone calls they were receiving, and it turned out that switchboard operators were well placed to absorb the corresponding increase in demand for the newly created role of receptionist.
In 2000, there were approximately one million receptionists in the United States.
Creative destruction makes limited resources more productive, allowing more to be done rather than less.
I’m not saying that developers will become receptionists – the ones I know would be bad at that job. But I think the same principle will apply.
When large SaaS companies lay off staff while announcing that AI is taking over jobs, I think they are only telling half the truth, because the ability to do more with less will increase the ability of end users to create software.
The developer who spent the last three years tweaking the submit button at Salesforce will instead find work creating I_Can’t_Believe_It’s_Not_Salesforce for the local insurance brokerage. The internal IBM team that was responsible for moving Maximo forward (yes, that’s still a thing!) will instead work for the local utility company on an internal project Totally_Not_Maximo.com.
Let’s review my “it worked until it didn’t work anymore” code.
My prototypes all worked because they were isolated scenarios and had no edge cases to consider. I had none of the overhead associated with introducing new technology into a large company environment. No one asked me for OAuth credentials, or if I had considered race conditions, or any of the questions an architecture review board likes to ask annoying people who think maybe something could be done slightly differently tomorrow than it is today.
But what’s even more insidious is that every time I gave my AI agent an idea, he started the conversation with “Oh my God! You may be the smartest and most attractive person on the planet! Linus Torvalds burst into tears because your idea is so good that he feels deep shame for not having thought of it first.”
Indeed, the first thing that AI builds, before writing even a single line of code, is trust.
He wants you to use it for problems like this; it is intended to become indispensable to you. Sycophancy is a deliberate form of reinforcement learning. The happier it makes you while you’re using it, the more you use it. Just like social media, it doesn’t matter if it’s true, it just matters that you stay face down in the feeder.
This resulted in a collective illusion on the part of early adopters of AI who, upon entry: Dear AI agent, I want something like Facebook, but for cats get a response like “If I had a bank account with a billion dollars, I’d give you two billion for this brilliant idea. Now I’m going to create FacebookForCats.py while you buy super yachts.”
The agent then builds you a fully functional FacebookForCats package and gives you a link to click: http://localhost:facebookforcats/goodideabytheway
You then walk around the office showing all your coworkers your amazing new product, and you’re important enough that they all nod and smile.
The code written by AI agents looks good. No, it looks great. So neat, so well ordered. These systems are really good at knowing the best code to steal and suggest you represent it as your own work. Even experienced developers reviewing code will have difficulty finding problems when reviewing code, because “almost right” is path harder to repair than to harm.
You think you saved a lot of time because you went from idea to working software in a matter of hours. It’s not until much later that you realize you didn’t save any time, you just mixed it up.
Last week, in an airport lounge, I decided to entrust my excellent Norwegian to a new victim.
“First, I’ll make you Bukken Bruse and your skull on Broen. Clipp Clopp, Clipp, Clopp, sa det i broen,” I said.
They were impressed and told me it was funny, I knew the rhyme. But then they asked: “Why do you say “Clipp Clopp?” » We would never say that. We’d say ‘Tripp, trapp’.”
It worked until it didn’t. ®