My experience with AI assisted coding showed me something I wasn’t prepared for after a lifetime in software development.

I’ve been writing code for a long, long time. I’m getting close to that point in my career where I’ve seen enough frameworks, languages, and methodologies to feel like I have a pretty good handle on things. But a recent experiment completely shattered that feeling. I stumbled into what you might call “vibe-coding,” and it showed me just how powerful AI assisted coding can be—in a way that’s both incredible and a little bit scary.

It all started with a project that had become a six-month-long headache.

The Wall: Six Months and Almost Nothing to Show

My work involves creating custom programming languages, and for a while now, I’ve been trying to integrate a powerful C++ library called libtorch into my latest language, a Lisp. For those unfamiliar, libtorch is the C++ engine that powers PyTorch, one of the most important libraries in the AI world. This isn’t a trivial task.

I even brought a trainee on board to help. But after six months, we had barely made a dent. The official documentation was sparse, and finding useful, real-world C++ examples was next to impossible. Most people use this library through its Python (PyTorch) interface, so the C++ backend is a bit of a black box. We were stuck, and honestly, I was ready to shelve the entire project. The progress was so slow it was almost painful.

A Three-Day Experiment in AI Assisted Coding

Then, I decided to try something different. I’ve been hearing about developers using AI models, but I was skeptical. How could an AI possibly understand the nuances of a custom programming language it has never seen, and correctly wrap a complex, poorly documented C++ library for it?

I set up an AI model to work in what you could call an “agentic mode.” Think of it less like asking a search engine for snippets and more like having a tireless junior partner who you can guide with high-level instructions.

The result? In three days, I accomplished what my trainee and I couldn’t do in six months.

And I don’t just mean I got a few functions working. I’m talking about a complete, functional wrapper for the most critical parts of the library. But it didn’t stop there. The AI also generated:

• Full documentation for the new implementation.
• A step-by-step tutorial on how to use it.
• Hundreds of example scripts to test every single function and ensure it all worked as expected.

The code compiles and runs perfectly on both macOS and Linux, with full support for GPUs. It just… works. Three days. I’m still struggling to wrap my head around it.

Beyond Prompts: What This New “Vibe-Coding” Felt Like

This experience wasn’t just about feeding prompts and getting code back. It felt more intuitive, like the AI understood the vibe of what I was trying to build. I was guiding the overall direction, the architecture, and the end goal, and it was filling in the massive gaps, navigating the cryptic library, and generating the entire supporting ecosystem. This is the future of AI assisted coding—not just as a syntax helper, but as a genuine collaborator.

Tools like GitHub Copilot have already shown us a glimpse of this, but this felt like a significant leap forward. It wasn’t just completing lines; it was completing entire concepts, from implementation to documentation.

Why I’m Worried About the Next Generation of Coders

As someone who has spent a lifetime learning the hard way—poring over documentation, debugging obscure errors, and building things from scratch—I have to admit, I’m worried.

This tool allowed me to bypass the very struggle that builds deep knowledge. The painstaking process of figuring out the libtorch API is what would typically forge an expert. But I skipped that entirely.

So, what happens to the next generation of developers? How will they learn the fundamentals when an AI can pave over all the difficult parts? There’s a real risk that we could create a generation of programmers who are great at directing AI but don’t have the foundational knowledge to build, debug, or innovate when the AI gets it wrong. It’s an incredible productivity boost, but what’s the long-term cost?

My little experiment with AI assisted coding was a success, but it left me with more questions than answers. It’s clear this technology is changing what it means to be a developer. The job might become less about writing lines of code and more about being a great architect, a great problem-solver, and a great guide for our new, incredibly powerful AI partners.

I’m still processing it, but one thing is for sure: the ground is shifting beneath our feet. For more technical details on the library at the heart of this, you can check out the official PyTorch C++ documentation. It might give you a sense of the complexity we were up against.