Yesterday, my kid sprawled on our couch, completely absorbed in her handheld console playing Pokémon. Looking at it, I felt a sudden urge to play Mario, just as the nostalgia kicked in.
It was pure nostalgia and fun, clearing those few initial levels as the muscle memory clicked in.
And that jump? It felt so real, I remembered how my friends and I used to physically pull the controller upwards, moving along with the flow, acting as if Mario would know we were trying to help him jump higher. It felt exactly as it did decades ago. Nothing jittered, nothing drifted, and nothing felt wrong. Zero input lag, pixel-perfect collisions.
Later in the day, I thought, “Let’s see if I can build this.” I wanted to recreate that magic, but I ended up feeling embarrassed instead.
So I sat powered on my Macbook Pro with 32 GB of RAM and high-speed internet and opened the latest Google Antigravity, my preferred AI coding IDE these days.
I typed:
You are a senior game developer. Build a playable, extensible 2D platformer with proper controls…
I hit Run. It generated files, assets, and lines of code.
Mario appeared… trapped inside a white box. The clouds looked like a Venn diagram from a bad LinkedIn post. The overall feel was pathetic.
AI Mario :)
And that’s when the irony landed. When I revisited the engineering history of that era, the year I was born, 1983, Shigeru Miyamoto and Gunpei Yokoi had just released Mario Bros. for arcades. Two years later, Super Mario Bros. arrived on the NES, designed by Miyamoto and engineered under Yokoi’s leadership, a defining moment created by a small Nintendo team working under extreme hardware constraints.
The entire game was just 40 kilobytes of storage, a game that even today feels relevant, maybe for me if not for Gen Z.
But let that sink in: 40 KB, much smaller than the screenshot posted above, even smaller than the favicon on most websites today.
To ship something timeless under those constraints, they couldn’t just be competent developers. They had to be technical minimalists with absolute clarity of thought.
- The Clouds? They are just recolored bushes.
- The map? Many small environmental tiles are reused in different combinations to create variety without spending extra ROM space.
- The Logic? They used assembly language. Every byte mattered, and every decision had consequences.
Now compare that to what my AI produced.
This model runs on data centers that draw enough power to light up a small city. It’s trained on millions of images of Mario scraped from the internet. And yet, Mario is trapped in a white div box. The clouds look like a Venn diagram drawn by a drunk mathematician.
It’s basically the difference between a parrot reciting Shakespeare and a poet writing it, one knows the pattern. The other understands the meaning.
Don’t Misunderstand Me
I use AI tools daily. Most of my boilerplate code is AI-generated, and 90% of my test cases are AI-generated. These tools have absolutely accelerated my work, and they belong in a modern engineer’s toolbox.
But watching a glitchy AI struggle with something a small team shipped flawlessly in 40 KB nearly 40 years ago was humbling.
Those engineers didn’t have autocomplete, they had graph paper and logic. They didn’t Google their way out of problems at 2 AM or ask Copilot to provide a solution by copying the error and pasting it in Copilot chat, they understood the hardware, the constraints, and the why behind every trade-off.
So yes, tools matter, they always have. But let’s stop pretending that scale alone replaces understanding, A continually expanding context window is not a substitute for architectural brilliance
A large language model can imitate output, it cannot replace the intentional, constrained genius of an engineer who knows why something works or how to think around those constraints at the base of the architecture
Conclusion
Perhaps we’re celebrating our tools a little too loudly, and I’ll admit, at times I’m guilty of that too. In the rush toward infinite scale, chasing a bigger context window, we’ve forgotten the quiet brilliance of constraint.
Use AI for the boilerplate. Let it accelerate the obvious. But when it comes to architecture and trade-offs, we should be the ones driving it, not the AI. Tools can help you move faster, they cannot help you think deeper.
When the most advanced AI models struggle to reproduce what a small team achieved in 40 KB four decades ago, it’s a reminder that the next real breakthrough won’t come from models trained on past patterns. It will come from engineers who understand why the code exists in the first place.
Mario is the king of games for me, and now not just because of nostalgia, but because it was built by people who mastered the fundamentals.
“Great engineering ages, no matter what the era is.”
Top comments (0)