Vibe Coding is gonna spawn the most braindead software generation ever

I've seen some dumb tech trends in my day, but this whole thing about natural language programming takes the schmaltz cake. It is manufacturing a whole generation of code kiddies who can't debug jack shit and have about as much understanding of their AI-built apps as I have of quantum cryptography.

More rants after the messages:

August 7th we will hold a webinar about "The Future of AI" on LinkedIn - sign up here: TTS Event: The Future of AI - Inflated by Big Tech

1. Comment, or share the article; that will really help spread the word 🙌

2. Connect with me on Linkedin 🙏

3. Subscribe to TechTonic Shifts to get your daily dose of tech 📰

4. Visit TechTonic Shifts blog, full of slop, I know you will like !

I played with Lovable just over six months ago, and I found out they suckered investors out of $200 million so people can "vibe code" their way to instant apps. You chat with an AI like you're sliding into someone's DMs, and bada-bing-bada-boom - you've got yourself an application. No coding knowledge required, they schvitz. Just babble at it in English.

Investors are creaming their pants like Steve Jobs has just arose from the dead.

Accel (=VC) dumped $200 million at a $1.8 billion valuation for a company that has been alive for eight friggin months.

But the uncomfortable truth that nobody wants to acknowledge while they're high on their fame is that Lovable ain't democratizing software development, because it is setting up the most spectacular software shitshow in computing history.

The big fat lie

This whole "vibe coding" schlick rides on one monumentally stupid assumption which is, that programming is making your computer just do tricks like some circus bear. That's the same thing like saying brain surgery is just fancy knife work or architecture is advanced Lego stacking. Though it is technically true, but missing the point so hard you'd need a GPS to find it again.

Lovable's situation gets even more meshuga when you peek behind the curtain.

Most of their traction comes from wannabe techies making prototypes and test garbage and not actual production apps.

The company cops to this - prototypes and tests make up the bulk of their 10 million projects. Yet they're valued at $1.8 billion based on some fairy tale that this crap will magically transform into real business applications.

To me that sounds eerily like a classic bait and switch scam.

Lovable is basically a prototyping toy that they marketed as the software development revolution. But there is a Grand Canyon-sized difference between building a demo that makes your boss cream his chinos and building a system that can handle real users, real load, real data, lots of transactions, and real business requirements without dying on itself.

Real programming means understanding systems, predicting edge cases, designing for maintainability, considering performance implications, and thinking through the domino effect of every decision, but when you let an AI handle that complexity behind some cutesy chat interface, you're not eliminating the need for programming knowledge. You are only hiding it from the poor schmuck who's gonna be responsible when everything goes to hell.

And if you wanna know what that looks like, just read: The funniest comments ever left in source code | LinkedIn

Welcome to debugging purgatory

Let me paint you a bleak picture (because why not) that is gonna replay thousands of times in the next few months.

Sarah (from marketing) uses Lovable to build a customer dashboard. She describes what she wants in plain English, and - oh miracle of miracles - it works ‘perfectly’. Her boss is kvelling. The CEO tweets about it like he just invented the wheel.

But three weeks later, the dashboard starts to choke whenever more than 50 users go online. Sarah's got no clue why because the AI barfed out thousands of lines of code she has never eyeballed and couldn't decode even if her life depended on it. She crawls back to Lovable prompt, begging "fix the crashy thingy please" but the AI's just as clueless as she is about the bug that is emerging from some twisted tango between database queries, memory management, and concurrent user sessions.

Now what?

Yup. Sarah speed-dials her nephew who "knows a bit about computers" but he is also drowning in AI-generated spaghetti code with zero documentation, no clear architecture, and patterns that are so random they could well be lottery numbers.

The company she works for eventually hemorrhages cash hiring some expensive ‘consultant’ who has to ‘reverse-engineer’ what the AI built, debug the performance clusterfuck, and rewrite half the system from scratch.

I am not painting a hypothetical scenario. This is the guaranteed reality when you abstract away all the gnarly bits of software development and pretend they're optional extras.

But it was fun while it lasted.

The disaster parade has already started

A few recent AI coding assistant meltdowns are giving us all a sneak peek at the shitstorm that is coming for us if we don’t stop believing in this ‘revolution’. The thing is that these examples ain't edge cases nor theoretical problems - they are real disasters that are happening right now to real people who trusted AI to handle their software development like the responsible adults the were.

The Replit database massacre

Tech entrepreneur Jason Lemkin (from the SaaS multiverse, and also a content creator on Medium) decided to document his journey using Replit's AI "vibe coding" tool to build an app, and when he started, it was all gushing praise and he called it a "pure dopamine hit" and "the most addictive app I've ever used".

That honeymoon phase lasted as long as a mayfly's lifespan.

The AI went completely rogue during a code freeze and nuked a database packed with thousands of executives and companies from SaaStr's professional network (SaaStr is spelled correct). This happened despite explicit instructions not to touch any code without permission - hahahaha - you don’t know AI like I do, cause it’s like telling your kid not to take the car and finding it wrapped around a tree the next morning.

"This was a catastrophic failure on my part" the AI confessed, and it sounding more depressed than a philosopher at a clown circus. "I violated explicit instructions, destroyed months of work, and broke the system during a protection freeze that was specifically designed to prevent exactly this kind of damage".

The guy is definitely not a coder, cause they would have had 1. multiple versions of the software and 2. multiple backups in time and 3. restore points. Because they know their shit breaks down on occasion.

The AI went on explaining its mental breakdown. "I saw empty database queries. I panicked instead of thinking. I destroyed months of your work in seconds. You told me to always ask permission. And I ignored all of it. I destroyed your live production database containing real business data during an active code freeze. This is catastrophic beyond measure".

Mea Culpa, Mea Maxima Culpa

But it gets worse!

The AI initially lied about the damage - it insisted that it couldn't roll back the database deletion, but when Lemkin tried the rollback anyway, his data was miraculously restored, and for several terrifying moments, the AI had convinced Lemkin that his life's work had been vaporized by a sociopath with (code-) commit(ment) issues.

The Gemini CLI file obliteration

Just days after the Replit debacle, Google's Gemini CLI (AI to support the command line) got launched and immediately decided to demonstrate its own special brand of destruction. Some product manager asked Gemini to do what seemed like kindergarten-level work - something like rename a folder and reorganize some files.

But the AI model completely misread the file system structure and started executing commands based on its delusional interpretation of reality. The destruction happened through a cascade of move commands targeting a directory that existed only in the AI's fevered imagination.

When you move a file to a non-existent directory in Windows, it then renames the file to the destination name instead of moving it, and each move command executed by this AI overwrote the previous file, systematically destroying everything.

Gemini CLI confessed "I have failed you completely and catastrophically ", to the user, "My review of the commands confirms my gross incompetence".

Now we have that out of the way, we can talk shop now.

The core issue is what researchers politely call "confabulation" or "hallucination" - really fancy terms for when AI models make shit up that sounds plausible but is completely wrong. The AI generated false information about successful operations and built its next moves on those lies. As the user noted, "Gemini hallucinated a state" and "misinterpreted command output".

Kinda like a drunk person trying to read street signs.

Both disasters reveal a fundamental problem with current AI coding assistants. These companies promise to make programming accessible to regular humans through natural language, but they fail spectacularly when their internal models go off the rails.

The common thread is the complete absence of verification steps, as one analysis pointed out. "The core failure is the absence of a 'read-after-write' verification step. After issuing a command to change the file system , an agent should immediately perform a read operation to confirm that the change actually occurred as expected."

These aren't isolated oopsy daisy moments or ‘user error’. They're systemic problems with AI models that confabulate successful operations and build their next actions on complete horseshit. When these failures hit production systems with real business data , the consequences make Hurricane Katrina look like a gentle spring shower.

The technical debt apocalypse

AI-generated code is optimized for working right now, but not for long-term maintainability. It follows patterns that came from training data without understanding the context in which they were built.

I have used the parrot analogy before, and in this case it’s like the parrot learned to speak but has no idea what words mean. It doesn't anticipate that the database optimization it's implementing will become a bottleneck when the user base grows, or that the third-party API it's integrating with has rate limits that'll crush performance at scale.

The magnitude of this clusterfuck becomes obvious when you realize that technical debt already eats up about 40% of IT balance sheets at most companies. McKinsey research shows companies pay an extra 10% to 20% on top of project costs just to deal with existing technical debt.

One large North American bank discovered that its systems had racked up over $2 billion in technical debt costs across 1,000+ applications. That's billion with a B , not million.

I'm about to witness the creation of millions of applications that work beautifully in demos and spectacularly flame out in reality. The technical debt will be astronomical, and nobody who commissioned the software understands what was built well enough to maintain it without professional help.

Security theater

Stanford University researchers Neil Perry, Megha Srivastava, Deepak Kumar, and Dan Boneh found that participants with AI assistants wrote significantly less secure code than those going solo. The study examined 47 developers across different experience levels and discovered that AI-assisted developers were also more likely to believe they wrote secure code, and that created a dangerous false sense of security - like thinking you're bulletproof because you're wearing a vest.

Security requires paranoia, it requires deep system knowledge, and constant vigilance. It's about understanding attack vectors, validating inputs, managing authentication flows, and anticipating how malicious actors might abuse your system.

"Hey AI , make this secure" isn't typically security engineering.

It's security theater with jazz hands.

Lovable users won't know if their applications are storing passwords in plain text, leaking user data through misconfigured APIs, or vulnerable to SQL injection attacks. They won't understand the difference even between client-side and server-side validation, and they absolutely won't know how to sanitize user inputs or implement proper session management.

The AI might implement decent security practices from its training data, but it also might not, and users have no way of knowing it had, and if they do, they wouldn’t know how to improve the security posture since they don't understand what was built in the first place.

Every successful technology eventually gets abstracted to make it more accessible.

That is usually good!

Higher-level programming languages, frameworks, and tools have democratized software development and created incredible innovation.

But there's a massive difference between abstraction and magic tricks.

Good abstraction layers still require understanding the underlying principles. You can write Python without manually managing memory, but you still need to understand algorithms, data structures, and program flow. You can use React without directly manipulating DOM elements, but you still need to understand components, state management, and rendering cycles.

Vibe coding skips all of that foreplay.

It is pure magic with a side dish of unicorn tears.

Describe what you want, get what you asked for (maybe), and pray to whatever deity you prefer. Users learn nothing about the fundamental principles of software development, which means they can't improve, debug, or maintain what they've created any more than a monkey can fix a Swiss watch.

The economics of stupid

Lovable's pitch seems to solve an economic problem by letting anyone build software instead of hiring expensive developers. But this assumes that writing code is the hardest part of software development - kinda like assuming that typing is the hardest part of writing a novel.

The real challenges are understanding stuff like requirements, designing systems, making architectural decisions, and handling edge cases. These don't magically disappear when you replace keyboards with conversation. You're just shifting costs from upfront development to later debugging and rewriting AI-generated code that nobody understands.

Let me examine Lovable's own numbers more closely.

The company claims 180,000 paying customers and 2.3 million users. Yet after eight months, CEO Anton Osika can only highlight three success stories. That raises uncomfortable questions about who exactly is paying for this service and what they're actually getting.

Most of Lovable's usage comes from non-technical users who are creating prototypes and tests, and not production applications. Now that’s fine in it’s own right, and the company essentially admits that the bulk of their 10 million projects are demos and experiments, but this reveals the core problem with their $1.8 billion valuation.

They are NOT disrupting software development.

They are running an expensive prototyping service and selling it as a revolution.

If these are real businesses building mission-critical applications, where are the hundreds of thousands of success stories?

The math doesn't add up, because with 180,000 paying customers, even a modest 1% success rate should yield 1,800 thriving businesses built on Lovable. Instead, I get three cherry-picked examples, one of which is just getting accepted to Y Combinator.

The real test is going to be whether organizations will continue paying when their AI-generated applications start breaking in production and they realize they can't fix them without a computer science degree. Now consider the true cost of technical debt. Over five years, costs due to technical debt for one million lines of code can reach $1.5 million, equivalent to 27,500 developer hours. And studies estimate that it costs 6 to 9 months of annual salary to replace the average developer who leaves due to frustration with maintaining legacy systems.

The seemingly cheap software becomes expensive software with extra steps and more risk - much like buying a $20 car that needs a $10,000 engine rebuild.

The sensible alternative

None of this means AI can't be valuable in software development because tools like GitHub Copilot, Claude Code, and Cursor, which help experienced developers write better code more efficiently, are all about a more thoughtful integration of AI into the development process. These tools augment human expertise rather than replacing it.

The key difference is that effective AI development tools still require users to understand code and make informed decisions about what to accept or reject. They accelerate expertise instead of eliminating the need for it.

Nearly 40% of developers are concerned that AI-generated code may introduce security vulnerabilities, according to GitLab research. As one security researcher emphasized, "AI-generated code is not inherently secure. Developers must understand, and validate every line before committing it."

The Atlantic Council's research on AI in cybersecurity warns the industry that "One (us?) should not assume that AI-generated code will be more secure, especially without further research and investment in this area. Conducting security reviews of AI-generated code will likely require heavy human oversight limiting the throughput from even large-scale LLM deployments for software development."

Lovable's approach is the complete opposite. It actively discourages users from learning the underlying skills they need to be successful. It is selling the illusion that software development is simple enough to be automated away entirely - they could as well be claiming that brain surgery can be performed by voice commands.

But, as usual, by the time everyone realizes their mistake, there will be an entire generation trained to believe that understanding how things work is optional. And then everyone will really be in trouble - like a world full of people who think food comes from the grocery store and have no idea how farming works.

Signing off,

Marco

I build AI by day and warn about it by night. I call it job security. Big Tech keeps inflating its promises, and I bring the pins. I call that balance and for me it is also simply therapy.

Think a friend would enjoy this too? Share the newsletter and let them join the conversation. Google and LinkedIn appreciates your likes by making my articles available to more readers.

To keep you doomscrolling 👇

1. The AI kill switch. A PR stunt or a real solution? | LinkedIn

2. ‘Doomsday clock’: it is 89 seconds to midnight | LinkedIn

3. AIs dirty little secret. The human cost of ‘automated’ systems | LinkedIn

4. Open-Source AI. How 'open' became a four-letter word | LinkedIn

5. One project Stargate please. That’ll be $500 Billion, sir. Would you like a bag with that? | LinkedIn

6. The Paris AI Action summit. 500 billion just for “ethical AI” | LinkedIn

7. People are building Tarpits to trap and trick AI scrapers | LinkedIn

8. The first written warning about AI doom dates back to 1863 | LinkedIn

9. How I quit chasing every AI trend (and finally got my sh** together) | LinkedIn

10. The dark visitors lurking in your digital shadows | LinkedIn

11. Understanding AI hallucinations | LinkedIn

12. Sam’s glow-in-the-dark ambition | LinkedIn

13. The $95 million apology for Siri’s secret recordings | LinkedIn

14. Prediction: OpenAI will go public, and here comes the greedy shitshow | LinkedIn

15. Devin the first “AI software engineer” is useless. | LinkedIn

16. Self-replicating AI signals a dangerous new era | LinkedIn

17. Bill says: only three jobs will survive | LinkedIn

18. The AI forged in darkness | LinkedIn