The Siren Song of New Tech
Every morning, I open Twitter to find my feed transformed into an AI wonderland. Demos of agents writing entire applications. Videos of developers “vibing” with their AI pair programmers. Posts celebrating the latest breakthrough in large language models. It’s intoxicating.
I know the feeling well. Just last month, I watched my team at Supervisible ship features at twice our normal speed, all thanks to AI assistance. The promise of “build faster, ship sooner, let AI handle it” isn’t just marketing—it’s real.
But here’s what those viral tweets don’t show: the afternoon I spent debugging after blindly accepting AI-suggested changes across dozens of files. The frustration when Cursor’s Agent mode aggressively rewrote our authentication flow. The head-scratching moments when Claude 3.7 suddenly generated inconsistent code compared to 3.5. Each breakthrough comes with its own set of gotchas, hidden just beneath the surface of those perfectly crafted demos.
The Reality Check
These aren’t just isolated incidents. Talk to any team building production software with the latest AI tools, and you’ll hear similar stories. Behind every slick demo lies a maze of complexity that no amount of prompt engineering seems to solve.
Even Andrej Karpathy—AI wizard and vibe coder extraordinaire captured this perfectly. He compared building modern web apps to assembling IKEA furniture, but without the courtesy of batteries included. No neat package, just a scattered collection of services demanding attention: authentication, storage, analytics, monitoring, and a dozen other pieces that never quite fit together as smoothly as the tutorial promised.
—recentlyThe irony? Karpathy admits spending three hours just configuring a basic Supabase setup with Vercel. When one of the world’s leading AI researchers struggles with dev tools, it tells you something about the state of our industry. “The second you stray from the getting started tutorial,” he notes, “you’re suddenly in the wilderness.”
This is the reality check we need: Building real applications isn’t about writing clever prompts or generating code. It’s about the unglamorous work of configuration, plumbing, and orchestration—the parts that no demo will ever show you.
Making Smart Choices
So how do we navigate this landscape? Dharmesh Shah, HubSpot’s founder, recently shared a perspective that resonated deeply:
“The world of A.I. moves really fast.” He points to Gemini 2.5 Pro, ChatGPT’s new image model, and the growing adoption of MCP.
But instead of chasing every new release, he offers a refreshingly practical framework:
“Experiment with the SOTA (State of the Art), Execute with the SOTP (State of the Practical).”
This deceptively simple advice cuts through the noise with a simple and effective explore/exploit algorithm: Yes, explore what’s new—try the latest models, test emerging frameworks. But when it comes to building products that deliver value today? Stick to battle-tested tools.
Or to put it another way: Choose boring technology. Not just because it’s proven, but because these battle-tested tools have generated millions of discussions, bug reports, and solutions across the internet—creating the very training data that makes AI assistants actually useful. When you pick a framework that’s been around for years, you’re not just getting reliability; you’re getting the collective wisdom of thousands of developers, now embedded in the LLMs helping you code.
At Supervisible, this approach has saved us countless headaches. While we experiment with AI features in controlled environments, our foundation runs on boring technology (Postgres, AWS, Cloudflare) that just works.
The Path forward
Everything is in motion. Each week brings new models, frameworks, and promises of simplified development. The pace isn’t slowing—it’s accelerating. Just look at your Twitter feed: another AI model, another deployment platform, another way to turn prompts into products.
Yet even the companies leading this revolution understand something crucial about innovation. Take Replit and Vercel. They’re working to abstract away infrastructure complexity, envisioning a world where apps materialize and provision themselves from prompts. But peek under the hood: v0.dev generates React components, Replit defaults to established frameworks. They’re building the future on battle-tested foundations.
This hints at a deeper truth about where software development is heading. Vibe coding isn’t just another trend—it’s a fundamental shift in how we build. Amjad from Replit captures it perfectly: clear thinking, precise articulation, and effective communication are becoming as crucial as coding itself. But don’t mistake this for the death of technical skills. If anything, understanding the fundamentals matters more than ever. The developers who will thrive are those who can grasp both the underlying principles, choose the right tools and the latest abstractions.