The AI Code Gap

I’ve heard hype about AI coding tools all year from every direction. After watching AWS re:Invent, it’s clear AI coding isn’t going anywhere. I still feel a disconnect between the dream and the reality, even though I can see how it could actually be achievable. The vague notion of how is the frustrating part—especially when I deal daily with AI solutions that fall over themselves doing basic things, while simultaneously pulling off insanely impressive complex things. It’s like moving incredibly fast and stubbing your toe every five seconds.

For the first time, the next leap in software development isn’t an abstraction layer. It’s inference. And inference generates code inconsistently. That’s why we can’t fully trust it yet—because it never does the same thing twice, and the results shift every time a new model or strategy drops.

The speed at which you (or anyone) can now write code is incredible. But my ability to understand code is still the same. Maybe a bit faster if I trust AI to do the comprehension for me, but still the bottleneck. So I spend most of my time trying to understand the writings of a madman with no grounding in reality, hurling almost-correct phrases at me. I’ve started calling this the AI code gap—the distance between what AI has generated and my ability to meaningfully understand or trust it enough to make changes.

This applies whether you can code or not. If you don’t know how to code, the gap only widens. I genuinely worry that a whole generation of new developers won’t learn by doing, because AI is so powerful that chasing errors or reading through mountains of code will feel pointless. And maybe that’s the real hurdle: the next generation of software development needs a trustworthy abstraction that sits above AI, allowing it to build systems we can rely on.

Codeless.

Not no-code, not low-code.

Codeless in the same way serverless doesn’t mean no servers—it means no servers to manage. I think we’re heading into an era where we create solutions without worrying about the code underneath, trusting that the underlying machine is handling it safely and consistently.

But honestly, software development hasn't evolved meaningfully in a long time. The number of missing “infrastructure primitives” for building an app in 2025 is baffling. Basic, universal features still require bespoke solutions. Why is rich text still not solved? Why am I still writing custom image resizing pipelines? Why are forms still hand-rolled nightmares? We need new infrastructure—trusted, AI-native building blocks that allow us to create apps without peering under the hood every 10 minutes.

That’s where the real exponential productivity unlock will come from. Maybe current tools like lovable are the first toe in the water, but the ocean is still ahead of us.

There will be an AI divide: people who harness AI effectively... and people who get gaslit by an overconfident, hallucinating junior that never learns from feedback and just keeps yeeting out slightly-wrong code until they give up.

We’re obviously not there yet—but 2026 looks like a year of serious change and disruption.