At the beginning of this month, GitHub and OpenAI launched the technical beta of Copilot. I got access a few days ago and installed it in my VS Code editor. To be honest, I was expecting a simple turbocharged autocomplete, something like IntelliSense that would guess the variable name.

But after a week of using it, I am genuinely scared and amazed in equal parts. My workflow for writing code has changed in a way I haven't experienced since we abandoned Notepad for modern IDEs.

Understanding context with OpenAI Codex

Under the hood, Copilot works with a model called Codex, a direct descendant of the GPT-3 architecture that we saw coming with GPT-2, but massively trained with billions of lines of public code from GitHub.

It doesn't just predict words; it predicts intentions. The other day I needed to write a Python script to parse some weird dates from an old log. Instead of going to StackOverflow, I simply wrote a comment and let Copilot do its thing:

# Función que recibe una fecha en formato "DD-MM-YYYY" y la convierte
# a un formato ISO 8601 estandarizado para la base de datos
def normalizar_fecha(fecha_texto):
    # [A PARTIR DE AQUÍ, TODO LO ESCRIBIÓ LA IA EN GRIS]
    import datetime
    try:
        fecha_obj = datetime.datetime.strptime(fecha_texto, "%d-%m-%Y")
        return fecha_obj.isoformat()
    except ValueError:
        return None

I hit the tab key and the code materialized. It imported the right library, used the strptime function with the exact mask for days and months, and added a try/except block to handle errors, something that I, in my laziness, probably would have omitted in the first version.

Reflection: From writers to code reviewers

The feeling of doing pair programming with a machine that has read all the code on the internet is surreal. Does this mean we are going to be out of a job? Not at all. Copilot is a monkey with a statistical machine gun; sometimes it suggests infinite loops or invents libraries that don't exist.

The role of the programmer is mutating. For the last fifty years, our main skill was memorizing syntax and translating human logic into text that the machine understood. With Copilot, that barrier disappears. Our new job is no longer "writing code", but "reading, validating, and orchestrating" the code the AI spits out. We have become the editors-in-chief of a brilliant but dangerously "imaginative" machine.