Alex de Pablos
AI

It's not about AI writing code

Alex de Pablos Lopez4 min

Today, AI writes code better than most developers I know. Faster, cleaner, and it doesn't get tired. I'm not saying this to provoke: it's what I see every day at work and in my own projects.

And still, I'm more and more convinced that the code is the least interesting part of everything that's happening. Let me tell you why.

My history with this (which might be yours too)

When ChatGPT went public, it was good for snippets. Context was limited and you needed a thousand tricks to get coherence out of anything longer than a chunk of code. But the snippets were good, and they already saved you the trip to Stack Overflow and half a morning of research. That's when I first thought: this is going to change everything. Though I'd be lying if I said I imagined where we'd be just three years later.

Since then I've gone through every phase, and if you're in my world they'll probably sound familiar. First, autocomplete and snippets inside the IDE. Then, a small agent that could already touch several files. After that, the CLI era in the terminal, approving changes one by one. And today, straight-up applications that manage not just your code but your whole environment: they open the browser to test the UI, connect wherever they need to, deploy to production. However far you give them permission, that's how far they'll go.

What letting go of the keyboard feels like

And yes, something gets lost along the way. Writing lines of code has a lot of romance to it: that flow state everyone talks about, you and your screen, building something from scratch. I miss it at times, I won't lie.

But practice rules. At first I reviewed every change under a magnifying glass. Over time, "accept, accept, accept" became the norm, until I reached a point that's uncomfortable to admit: I review the code because it's the thing to do and to feel good about it, knowing that most of the time what I'm reviewing is better than what I would have written.

The little monsters

Does that mean you can check out? Not yet. If you don't give a project attention (or care), it turns into what I call a little monster: every single answer from the model is fine, it compiles, it works... but the whole grows without coherence. Unnecessary abstractions, utilities that already existed written again, and the classic trap: you start fixing a bug, get into the back-and-forth with the agent and, by the time it's fixed, you have two thousand more lines of code than yesterday.

Why does it happen? Because each request to the agent is, in quotes, independent. Managing the context of a large application is hard for them. Someone has to hold the vision of the whole, set the guardrails, say no to solutions that work but don't fit. For now, that someone is you.

And notice I say "for now", because this is also changing in front of our eyes: the latest models manage context better, repeat themselves less, and are even starting to simplify instead of just adding. It happens less and less. The direction is clear.

What's actually changing

That's why I say the code is the least of it. What's transforming is the whole process: how we work, how we create, even how we learn and how we think.

Judgment matters more, not less. When generating a solution takes seconds, the scarce skill is telling the good one from the one that just looks good. The bottleneck is no longer writing: it's deciding.

Work looks less like typing and more like directing. Clear goals, guardrails, verification stages, review, course correction. I already write far more text than code, always. And it goes further: I use the keyboard less and my voice more, whole sessions talking to the computer, dumping my thoughts as they come so it can sort them out. LLMs are excellent at that, and they'll only get better (even Karpathy was saying it just now).

And knowledge cannot be delegated. Another Karpathy idea that clicks with me: you can delegate the thinking, let another intelligence think and execute for you. What you can't delegate is the knowledge: knowing what is being done and why. Maybe mastering the how is no longer required. But the what and the why, those you can't let go of.

So how long does "for now" last?

Honestly, I don't know. No crystal ball here. What I do know is that learning to build from this new place, with its wins and its messes, is what this stage is about for me. And whatever I figure out, I'll keep leaving it here.