“After 12 months of evangelising AI-assisted development, I have a confession: I’ve been wrong. Not partially wrong, not ‘it depends’ wrong. Fundamentally wrong. Yesterday I uninstalled every AI coding tool I had.”
This was the first line of a post I made on LinkedIn last week.
It talked through my reasoning for abandoning agentic development and “vibe coding” after many months as a true believer who publicly advocated for AI at every opportunity. It was a big change in tone from my previous content, and it gathered a lot of attention.
After watching the post gather 40,000 views, I trawled through dozens of comments all saying versions of the same thing. “Congratulations,” these comments from highly-qualified software engineers said, “on finally seeing the light and leaving the hype behind. It was brave of you to own up.”
It was not the response I was expecting. But it was a response that told me a lot about the state of the software industry at this fascinating point in its history.
The post was an April Fools joke, of course. I wasn’t trying to hide it: the whole post was written by AI and ran in direct and violent contradiction to everything I’d been saying up to then. An astute observer would even have noticed the first letter of each paragraph in the post: A, P, R, I, L, F, O, O, L, and S.
Get the best of Startup Daily straight to your inbox
Want to know the latest in startup news? Subscribe to our daily news and analysis coverage on what’s happening to ANZ startups, investors and the broader ecosystem. And best of all, it’s FREE!
By continuing, you agree to our Terms & Conditions and Privacy Policy.
More to the point, I believed the premise was self-evidently absurd. There is a great deal of hype about AI from self-interested parties, shills, grifters, and hangers-on. It irritates me no end.
But even a stopped clock is right twice a day: every once in a while, the hype is justified. This is such a case.
As a software engineer of over 20 years, my experience of agentic coding tools like Claude Code has been nothing short of transformative. It has allowed me to build sophisticated software massively faster, and at stunningly good quality.
Things that would have taken me weeks or months to build now take days or sometimes even hours. This is not “AI slop”, either. It is production-grade software, carefully architected and fully instrumented.
The PC (Pre Claude) and AC eras
There was before Claude Code, and there was after.
Many other software developers I talk to feel exactly the same way: they are convinced not because they read breathless articles in the media and swallowed the hype, but because of their lived experience as practitioners, day after day.
What we all see is a change to the practice of software development that dwarfs anything wrought by the web, cloud, or mobile revolutions.
My April Fools post revealed that many intelligent, credentialled people have a very different view of AI-assisted software development. It forced me to ask myself: what is it about their experience with AI that has led them to such a different conclusion?
When I look at the comments affirming my volte-face, I see two threads of reasoning: first, that AI produces bad code that is hard to debug and maintain, thereby counteracting any perceived productivity gain; and second, that by blithely handing control to AI agents we are diminishing our own skills and agency as humans—to the long term detriment of both ourselves and the software we have built.
The observations are not without merit, but to therefore conclude that one should return wholly or in part to the not-long-ago days of writing code by hand is incorrect.
Dealing with GIGO
The phrase “Garbage In, Garbage Out” (GIGO) in computing dates all the way back to the 1950s. It refers to the fact that when faced with poor-quality or incorrect inputs, it is inevitable that computers will produce poor-quality or incorrect outputs.
Indeed, Charles Babbage, the 19th-century father of computing, recalled:
“On two occasions, I have been asked, ‘Pray, Mr. Babbage, if you put into the machine wrong figures, will the right answers come out?’ I am not able to rightly apprehend the kind of confusion of ideas that could provoke such a question.”
With AI, GIGO has been given a new pair of running shoes.
Careless or inexperienced builders can ship software consisting of tens or hundreds of thousands of lines of code without understanding how it’s built or what it does.
Prompts lacking in specificity, nuance, and architectural thought will lead directly to poorly-built software.
Building with AI
The issue is not that the AI tools are unequal to the task of building good products, but that if you don’t clearly specify what you want done and how you want it built, you should be unsurprised when the result is garbage. And indeed, what kind of confusion could provoke anybody to believe otherwise?
Where the skeptics get it right is that building good software products with AI takes a great deal of skill, care, and cognitive effort. Software engineers are not going anywhere, at least not for a while.
Using AI tools effectively also requires the acquisition of new skills and a willingness to unlearn and abandon long-cherished beliefs and aspects of our identity as developers.
Where I disagree with them is that there is any alternative to embracing AI in the building of software, and embracing it fully.
AI has changed software development profoundly: of that we have evidence and no doubt. It is no longer valid to ask whether we should use this technology.
For those who have tried using AI agents and found them to be lacking, the only question they should ask themselves is: what do they need to be doing differently?
