The Rumor Mill Around GitHub's AI Pair Programmer Is In Overdrive
Developers are swapping screenshots of a mysterious GitHub tool that autocompletes whole functions from a comment, and OpenAI's Codex looks like the engine.
I’ve lost count of how many screenshots have crossed my timeline this week of some GitHub tool finishing an entire function from nothing but a comment or a function signature. Nobody I know has an official name for it yet, but the pattern in the demos is consistent enough that it’s clearly one project making the rounds through a limited preview, and the people showing it off are, understandably, having a lot of fun with it.
The engine underneath appears to be Codex, a version of GPT-3 that OpenAI has fine-tuned specifically on public source code. That’s a meaningfully different animal from the GPT-3 most people have played with for text generation. Feed a general-purpose language model a docstring and you might get plausible-sounding prose back. Feed Codex a docstring or a half-written function signature and, per what I’m seeing in these clips, you get working code — sometimes eerily on-target, sometimes comically wrong, but often close enough to save real typing.
If this does turn into a shipped product, it would represent a genuine shift in how a lot of us write code day to day. Autocomplete has been creeping toward “finish my thought” for years — think snippets, then IntelliSense-style type inference, then the small predictive suggestions some editors already offer. A model trained on the sum of public GitHub code taking that to “write my function” is a different order of magnitude.
The questions nobody’s answered yet
Two things keep coming up whenever this tool gets discussed, and neither has a satisfying answer right now.
The first is licensing. Codex was reportedly trained on public repositories, which cover every license under the sun — permissive, copyleft, and plenty of code with no explicit license at all. If a suggestion is a near-verbatim reproduction of someone’s GPL-licensed function, does the person who accepts that suggestion inherit any obligations? Nobody outside the small preview group seems to have a clear answer, and I doubt GitHub does either at this stage.
The second is accuracy, or really, trust calibration. A tool that’s right 80% of the time and confidently wrong the other 20% is arguably more dangerous than a tool that’s obviously unreliable, because developers start pattern-matching “this looks right” instead of actually reading the code they just accepted. Anyone who’s shipped a subtle off-by-one bug because a suggestion looked plausible knows exactly what I mean.
None of this is a knock on the idea — I think an AI that can draft boilerplate, guess at obvious implementations, and free up your attention for the actually hard 20% of a problem is a genuinely great use of the technology. But the gap between “cool demo” and “tool developers can trust with their commit history” is usually wider than it looks from the outside. I’ll be watching closely for whatever GitHub announces next, because if this lives up to the screenshots, it’s going to change a lot of muscle memory in a hurry.