AI-titude

Human in the loop

I was at first firmly opposed to using AI to write code, for a simple reason: humans still wrote better code than it did. That is no longer true. Apart from a few very specific cases, AI now writes better code than a developer, and much faster.

Human know-how does not disappear; it moves. It now lies in architecture, in how the work is organised, and in securing the code that AI produces.

Command of languages, code structures and development techniques is still necessary. A new skill comes on top of it: knowing how to design and properly use AI harnesses, that is, the framework of instructions, tools and checks within which AI works. A harness makes it possible to express a developer’s method and techniques, then to automate them.

That is why I now focus on architecture, audit and consulting. AI supports me in that work, but know-how is what decides: what to build, how to frame it, and what can be accepted from what it produces.

I still produce the code that implements my architectures. But it is now mostly generated from specifications: architecture plans, and contracts, meaning the guarantees and constraints the code must meet, turned into tests and checks.

The code I still write entirely by hand is for my personal projects: hobby and side projects unrelated to my professional work, usually in exotic or highly technical languages. That is where I keep exploring, for the sheer pleasure of code.