GitHub Copilot Wants to Finish Your Sentences

GitHub just dropped something that’s going to change how a lot of us write code, at least for the rest of the afternoon while everyone tries it out. It’s called Copilot, built with OpenAI, and it’s an AI pair programmer that lives inside your editor and suggests whole lines or even entire functions as you type.

Under the hood is Codex, a descendant of GPT-3 that’s been fine-tuned specifically on public source code. That distinction matters. GPT-3 is a general-purpose language model that happens to be decent at code because code showed up in its training data. Codex is the inverse bet: take that same architecture and specialize it hard on code, so it actually understands idiom, structure, and the shape of a function before you’ve finished naming it.

Right now it’s a free technical preview, and it only works as a Visual Studio Code extension. You install it, sign up for the waitlist (or already have access, if you’re lucky), and it starts making inline suggestions as you write. Type a function signature and a comment describing what it should do, and Copilot will often just… write the function. It’s not always right, and you still need to read what it gives you, but the hit rate on boilerplate and common patterns is startling.

At launch it works best with Python, JavaScript, TypeScript, Ruby, and Go — which tracks, since those are exactly the languages with the most public code on GitHub to train on. If you’re writing something more obscure, don’t expect the same magic.

The interesting part is the business model, or lack of one

For now, Copilot is free. GitHub hasn’t said how long that lasts, but I’d bet real money this doesn’t stay free forever — training and running a model like this isn’t cheap, and GitHub is owned by Microsoft, a company that generally figures out how to charge for things eventually. Enjoy the free preview while it lasts.

There’s also going to be a fight about what this means for open-source licensing. Codex was trained on public repositories, which presumably includes plenty of code under licenses that say something about attribution or reuse. GitHub’s position seems to be that suggestions are transformative enough not to trigger those obligations, but I expect that argument to get tested, possibly in court, before this is over.

Practically speaking, though, this feels less like a novelty and more like the first real glimpse of autocomplete growing into something closer to a junior developer sitting next to you. It won’t replace understanding your own code — you still have to know what you’re asking for and verify what you get back — but for scaffolding, repetitive patterns, and shaking off the blank-page problem, it’s genuinely useful. Worth installing this week just to see where the line between “helpful” and “eerie” actually sits for you.