I saw that post, neat stuff. We made an attempt to develop something similar 4 years ago and take it to YC, it simply wasn't good enough often enough because our training data (Stack Overflow posts) was garbage and models were weaker back then. I figured it would take about 5 years for it to really be useful given the technology trajectory, and here we are.

I'll note that we weren't trying to build "code auto-complete" but instead a automated "rubber duck debugger" which would function enough like another contextually-ignorant but intelligent programmer that you could explain your issues to and illuminate the solution yourself. But we did a poor job of cleaning the data and we found that English questions started returning Python code blocks, sometimes contextually relevant. It was neat. This GitHub/OpenAI project is neater.

I would be curious what the cost of developing and running this model is though.