The Awakening
From discovery to stepping back.
Late 2022 — the first shock (GPT-3.5)
On November 30, 2022, ChatGPT launches. I try it right away and stay on it for hours. First intensive session with GPT-3.5: generating music, text, bits of code, testing ideas in every direction. At that point, it’s mostly curiosity: is this tech actually useful for coding? Short answer: yes, but not without guardrails.
The first scripts are sometimes impressive, but the tool also invents APIs, nonexistent imports, and anti-patterns. The lesson comes fast: AI can produce, but you have to verify. Human oversight stays central.
March 2023 — the GPT-4 leap and the discovery of Cursor
On March 14, 2023, GPT-4 comes out, and I feel a real jump. From the first tests, the model understands better, reasons further, and writes more coherent code. It’s the first time I tell myself: “AI is starting to code almost like a colleague.”
Shortly after, I discover Cursor, designed around AI from the start, not as an add-on extension. I adopt it, test autocompletion (AI proposing the next part of your code), and experiment with .cursorrules to define global rules in my projects.
That starts a long exploration: context becomes the key to quality.
2023 — auto-completion and first rules (Copilot, Cursor)
Autocompletion has existed for years in code editors: type the beginning of a word, the tool suggests the rest. Before AI, it was mostly basic keyword completion from your own code. With AI, it changes scale: it anticipates what you’re writing, proposes entire functions, and adapts to project context. It’s continuous help, useful but demanding: quality depends on the context you provide (variable and function names, comments, pseudo-code), not only on the model.
From March 2023, Cursor introduces .cursorrules at project root: a simple, persistent way to add global rules. Then in December 2023, Cursor moves to a richer setup with the .cursor/rules directory, letting you organize rules hierarchically. That change matters: rules can target specific folders, with a level of granularity that changes how you manage context.
In that setup, I build my first .cursorrules files and gradually push the idea of context hierarchy. The core idea becomes clear: the context you give AI is the determining factor.
2024 — Maturation
In 2024, out of caution, I don’t use AI directly on the codebases of professional projects. I still use it on the side: snippets, brainstorming, and especially reformulation.
I also observe daily life in large companies. I keep seeing the ceremonies and rituals that became standard for team synchronization. I step back from that machinery: it’s necessary, but the “coordination cost” of keeping everyone aligned is very real.