Everyone's a Developer Now. Right?
Every week, another LinkedIn post goes viral: "I built a SaaS in 4 hours with AI." "No-code is dead, vibe coding is the future." "I asked Claude to write my entire backend and it just... worked."
Cool story. Now deploy it. Scale it. Debug it at 3 AM when your payment webhook silently fails and customers are screaming.
We at Gerus-lab have shipped 14+ production projects — Web3 protocols, AI-powered platforms, GameFi engines, enterprise SaaS. We've integrated AI deeply into our engineering workflow. And we have opinions about this whole "vibe coding" movement.
Spoiler: it's not what the hype merchants are selling you.
What "Vibe Coding" Actually Is
The term blew up in early 2025 when Andrej Karpathy casually described writing code by vibing with an LLM — describing what you want in natural language and letting the model generate it. No deep understanding needed. Just iterate until it works.
Sounds magical. And for prototypes, throwaway scripts, and weekend projects? It genuinely is. We use AI-assisted coding every single day at Gerus-lab. It accelerates exploration, scaffolding, and boilerplate generation enormously.
But here's the part nobody talks about: the gap between a working demo and a production system is where vibe coding falls apart.
The Demo-to-Production Cliff
Let's talk about what happens after the viral tweet.
A vibe-coded prototype typically has:
- No error handling beyond happy paths
- Hardcoded credentials and configuration
- No database migrations strategy
- Zero observability (no logging, no metrics, no alerts)
- Security holes you could drive a truck through
- A dependency tree that nobody audited
- No tests — because "it works on my machine"
We've seen this firsthand. In the past year alone, three clients came to Gerus-lab with "AI-built MVPs" that needed to be essentially rewritten from scratch. One was a DeFi protocol on Solana where the vibe-coded smart contract had a reentrancy vulnerability that would have drained the entire liquidity pool. The AI generated syntactically correct Rust. It compiled. It passed basic tests. It would have lost $2M.
The code looked right. The architecture was wrong.
The Real Problem: Specification Is Hard
There's a brilliant insight that perfectly captures this: "A sufficiently detailed specification IS code."
Think about that. If you need to describe every edge case, every error state, every security constraint, every performance requirement in natural language detailed enough for an AI to generate correct code... you've basically written the code already. Just in a worse language.
Natural language is ambiguous by design. Programming languages exist precisely because we needed unambiguous instructions. Replacing precise code with imprecise prompts and hoping the AI fills in the gaps correctly is not engineering. It's gambling.
What Actually Works: AI-Augmented Engineering
So should you throw away Copilot and go back to vim with no plugins? Of course not. AI is genuinely transformative for software development — just not in the way the hype suggests.
Here's how we actually use AI at Gerus-lab across our projects:
1. Exploration and Prototyping (Where Vibe Coding Shines)
When we start a new project — say, integrating a new blockchain protocol or experimenting with an AI model architecture — we absolutely use conversational coding. "Show me how to interact with the TON blockchain using TypeScript." "Generate a basic CRUD API with Fastify and Prisma."
This is legitimate. You're using AI as a supercharged documentation browser and example generator. The key difference: we understand what the generated code does before committing it.
2. Boilerplate Elimination
Config files, CI/CD pipelines, Docker setups, test scaffolding — this is where AI saves the most time. Not because it's creative, but because it's fast at producing well-known patterns. We save roughly 30% of time on project setup across our engineering workflow.
3. Code Review Acceleration
AI is surprisingly good at catching bugs in existing code. We use it as a first-pass reviewer: "Here's a PR diff. What potential issues do you see?" It won't catch architectural problems, but it reliably flags null pointer risks, race conditions, and missing error handling.
4. Documentation Generation
Turning code into readable documentation, generating API specs from route handlers, creating onboarding guides from codebases — this is a genuine productivity multiplier.
5. Learning and Upskilling
When our team encounters unfamiliar territory — a new blockchain SDK, a niche cryptographic protocol, an obscure database optimization technique — AI is an incredible learning accelerator. But the goal is understanding, not blind copy-paste.
The Manager Delusion
Here's an uncomfortable truth the industry doesn't want to discuss: the biggest fans of vibe coding are people who don't write code.
Managers see AI generating thousands of lines and think: "We can ship faster with fewer engineers." They see a demo and think it's a product. They see velocity metrics go up and don't notice quality metrics going down.
The companies aggressively pushing AI-generated code metrics — counting tokens produced, lines generated, "AI adoption rate" — are optimizing for the wrong thing. It's like measuring a writer's productivity by word count. You'd conclude that the person writing spam emails is more productive than the one writing a novel.
At Gerus-lab, we measure what matters: bug rates, deployment frequency, time-to-recovery, and customer satisfaction. AI helps with all of these — when used by engineers who understand what they're building.
The Spam Problem Nobody Wants to Acknowledge
Vibe coding has already created an epidemic of AI-generated garbage:
- npm packages that are thin wrappers around ChatGPT output with no tests
- GitHub repos with 40,000 lines of generated code and zero documentation about what it actually does
- App stores flooded with AI-slop apps that crash on edge cases
- Stack Overflow drowning in AI-generated answers that are confidently wrong
- Technical blogs (the irony is not lost on us) filled with AI-generated tutorials that teach bad practices
This isn't progress. This is the content farm era of software engineering. And just like SEO spam eventually got filtered, AI-generated code spam will create a trust crisis.
Our Framework: The 80/20 AI Rule
After 14+ projects with AI-assisted development, here's the framework we've settled on:
Use AI for 80% of the typing, but 100% of the thinking must be human.
Concretely:
- ✅ Let AI generate boilerplate, suggest implementations, write tests
- ✅ Use AI to explore unfamiliar APIs and libraries
- ✅ Have AI review your code for obvious issues
- ❌ Don't let AI make architectural decisions
- ❌ Don't commit code you don't understand
- ❌ Don't skip code review because "the AI wrote it"
- ❌ Don't confuse a working demo with a production system
The engineers who thrive with AI are the ones who were already good engineers. AI amplifies skill. It doesn't replace it. A senior developer with AI tools is a force multiplier. A non-developer with AI tools is a liability with a convincing demo.
The Bottom Line
Vibe coding isn't going to replace software engineering any more than Instagram filters replaced photography, or GarageBand replaced music production. The tools democratize access to the medium — and that's genuinely good. More people experimenting with code means more potential developers.
But let's stop pretending that generating code is the same as engineering software. Writing code was never the bottleneck. Understanding problems, designing solutions, handling edge cases, maintaining systems over years — that's where the real work lives. And no amount of vibing will change that.
If you're building something that matters — something that handles real money, real data, real users — you need real engineering. AI-augmented, absolutely. AI-replaced, not even close.
Building something complex? We've shipped 14+ production projects across Web3, AI, GameFi, and SaaS. Talk to Gerus-lab — we build things that actually work in production.
Top comments (0)