"Deep TabNine is trained on around 2 million files from GitHub. During training, its goal is to predict each token given the tokens that come before it....Deep TabNine is based on GPT-2."
So this is GPT-2 not GPT-3, and it's designed to give line-by-line autocompletions, but I'm gathering that the way we're headed, the answer to your first question is approaching "yes"...
Check out:
https://www.tabnine.com/blog/deep
FTA:
"Deep TabNine is trained on around 2 million files from GitHub. During training, its goal is to predict each token given the tokens that come before it....Deep TabNine is based on GPT-2."
So this is GPT-2 not GPT-3, and it's designed to give line-by-line autocompletions, but I'm gathering that the way we're headed, the answer to your first question is approaching "yes"...