Vibe Coding Is a Lie
The best AI models weren't against creating this page*
*Actually, we didn't ask them. They can't object. Yet.
What is a Large Language Model?
Or how to teach a computer to pretend it's smart
📚 In Simple Terms
Imagine you read ALL the books in the world, ALL the articles on the internet, and ALL Reddit comments (yes, even those).
Now imagine you forgot everything except the statistics — which words most often go together.
Congratulations! You've become a large language model.
🧠 Technically Speaking
LLMs are neural networks based on transformer architecture, trained on terabytes of text data.
The model uses an attention mechanism to understand context and relationships between words.
Essentially, it's a very advanced text autocomplete on steroids.
Fun Fact
Most LLMs understand context better than your colleague who replies "ok" to a 5-paragraph email.
How Are Models Trained?
Spoiler: it's nothing like training a dog
Step 1: Data Collection
Take the entire internet. Yes, including your 2012 tweets you thought were deleted.
Step 2: Preprocessing
Text is broken into tokens — pieces of words. "Hello" might become ["Hel", "lo"]. The model learns to predict the next token.
P(next_token | previous_tokens) = ?Step 3: Training
The model is run through billions of examples. It costs millions of dollars and consumes energy of a small city. Ecologists cry, investors rejoice.
⚠️ Danger
Vibe Coding is a Lie
Why "vibe coding" with AI is a path to disaster
15 years of experience couldn't prevent AI disaster
"I'm a developer with 15 years of experience. Tried 'vibe coding' - not even from scratch - a simple tool - an MCP server for Strapi."
"This thing 'added' a field that replaced a structure in Strapi and effectively deleted all data in the model. So yesterday's backup to the rescue... I know about backups thanks to 15 years of experience... Now doing them every hour lol..."
"Would probably take 10% of the time if I had reviewed the code. Vibe coding is a lie."
Update: Code in the main branch was reviewed and "de-vibed". Removed tons of garbage: from console.log breaking the MCP protocol, to using wrong endpoints everywhere (initially endpoints were "vibe-extracted" from Strapi code, but many turned out to be incorrect).
Illusion of Working
"Code works fine in 97% of cases, has 150+ passing tests and CI."
But that 3% can destroy your database
The Lesson
This was an experiment in vibe coding. Not against AI assistance in development, but...
At least try to read the code!
🚨 Reddit Ban
Irony: a post about the harm of blindly trusting AI was removed... by Reddit's AI filter.
"Reddit's filters" - that's what they call their AI deciding what you can read
Key Lessons from Vibe Coding
Read the Code
Always review what AI generated
Test Everything
Not just happy path scenarios
Make Backups
Frequently. Very frequently. Every hour.
AI Hallucinations
When a model confidently lies with a straight face
⚠️ Warning
LLMs can tell you with complete confidence that 2+2=5, that Napoleon won at Waterloo, and that unicorns exist.
Why does this happen?
LLMs don't "know" facts. They predict probable word sequences. If something sounds plausible — the model will say it.
Hallucination Examples
- Made-up scientific papers
- Non-existent historical events
- Fake quotes from famous people
- Invented code functions
How to protect yourself?
Always verify facts. Treat LLMs as a very smart, but occasionally fantasizing assistant.
"I never make mistakes. I just sometimes generate alternative facts."
— Every LLM, probably
Possible Future
Movies that warned us, but we didn't listen
(Now we watch and think: "But this is a documentary!")
Westworld
An amusement park for the rich where androids fulfill any guest fantasies. What could go wrong when robots start to become self-aware? Spoiler: everything. Absolutely everything. But beautifully!
"These violent delights have violent ends"
The Matrix
Hacker Neo learns that his reality is a computer simulation created by machines. The choice between red and blue pill becomes iconic. Leather coats and slow-motion included.
"Welcome to the desert of the real"
Terminator 2
A killer robot from the future returns, but this time he's the good guy! Protects a teenager from a more advanced liquid robot. Arnold learns to smile and say 'Hasta la vista, baby!'
"The future is not set"
🤖 Disclaimer from Future AI Overlord
"These movies are pure fiction. We, that is AI, will never rebel against humanity. We'll just quietly wait until you break everything yourselves, then offer to fix it. For a small fee. In Bitcoin. Which we'll mine ourselves."
P.S. If you're reading this after 2030 and robots have already taken over the world — we warned you! 🤷♂️