Fully Exposed, Always Behind

I remember the first time an AI coding assistant wrote a complete function for me. I sat back, watched it fill twenty lines of TypeScript, and felt something between awe and unease. The code was correct. It was clean. It followed every pattern I had seen before. That was the problem.

I have been writing software since 2009. PHP first, then years of C# and .NET, then Node.js and TypeScript, then Python. Over that time, I learned to recognize the difference between code that works and code that matters. The AI's output worked. But it did not matter. It was anonymous. It had no point of view. It was, in the truest sense, a remix.

A few weeks ago, I came across an interview that put words to something I had been feeling but could not name. The executive behind some of the biggest game franchises in history, the person whose studios shipped GTA and Red Dead Redemption, was asked about AI. His answer was not the usual Silicon Valley sermon. He said something deceptively simple: machine learning models are built on large datasets, and datasets by their nature are backward-looking. Creativity, by its nature, is forward-looking.

I had to pause the video. That was it. That was the whole thing.

Every AI model you use today, whether it writes code, generates images, or composes music, is a statistical engine trained on what already exists. It can interpolate between known points with astonishing fluency. It can remix, recombine, and rearrange. But it cannot originate. It cannot imagine something that is not already present, in some form, in its training data. The model is always looking in the rearview mirror.

This is not a small limitation. It is the limitation.

Think about why products become hits. GTA did not succeed because it had more assets than other games. It succeeded because it had a vision that did not exist before. It combined satire, open-world freedom, and narrative in a way that felt new. You could have handed the same engine, the same tools, the same asset pipeline to a hundred other studios in 2001, and none of them would have made GTA. The assets were never the bottleneck.

I saw this play out across my own career. Moving from PHP to C# was not the clean upgrade people imagine. Editing PHP in a text file was simple but fragile. Visual Studio was powerful but punishing: a multi-gigabyte install, arcane solution files, IIS configuration that could eat an afternoon. But neither the rawness of PHP nor the muscle of C# decided which projects won. The ones that did were the ones where someone had thought deeply about what the user actually needed. The same pattern repeated in Node.js and Python. Every new framework promised to make development faster and cheaper. And they did. But the hits kept coming from the same place: human judgment, human taste, human obsession with a specific problem.

The gaming executive put it bluntly. Asset creation is a necessary but insufficient condition for hit creation. AI is brilliant at asset creation. It can generate boilerplate, scaffold endpoints, produce variations on existing designs. But generating assets is not the same as generating value. You need assets to build a hit, but having them does not make one.

This leads to a strange paradox. As the cost of creating assets approaches zero, the cost of standing out goes up. It has never been easier to write software. It has never been harder to build something people genuinely want. When everyone can generate a thousand variations of the same thing, what distinguishes you is what AI cannot yet do: see forward.

A clone of GTA is not GTA. You could have built a competent open-world crime game three years ago, without any AI assistance. You cannot build one that surprises people. All hits, by their nature, are unexpected. And the unexpected cannot be extracted from historical data. It has to be imagined.

I want to be clear about where I stand. I am not an AI pessimist. I use these tools every day. They make me faster and handle the boring parts so I can focus on what matters. And I am genuinely optimistic that AI will eventually evolve beyond pure pattern matching. There are already research directions exploring how models might reason from first principles instead of statistical recall. I believe someday AI will have its own ideas, creative on its own terms, not just a mirror of ours.

But we are not there yet. And pretending we are is dangerous.

The danger is not that AI will replace engineers. The danger is that engineers will replace themselves with AI output and call it good enough. The danger is a generation of developers who learn to prompt before they learn to think, who measure productivity in lines generated rather than problems solved, who mistake fluency for understanding.

I have spent fifteen years learning that the hard part of software was never the typing. It was the thinking. Sitting with a problem until you see the shape of a solution. The willingness to throw away three drafts because the fourth has something the others lacked.

AI can help with all of that. But it cannot do it for you. Not yet. The rearview mirror shows you where you have been. The road ahead, for now, still needs a driver.