Great question!

TabNine is a different approach that uses far less semantic information than Kite. It really shines when you're talking about syntactic repetition, e.g. if you check out the screenshots on their homepage. They have a page about semantic completions, but the semantics are very shallow -- basically what attributes are on the instance you're accessing.

In contrast, we've spent ~50 eng-years semantically indexing all the code on Github, building statistical type inference, and rich statistical models that use this semantic information in a very deep way. The result is that Kite can help more often, in ways that reflect a deep understanding of the semantics of the code you're writing.

Not to mention by pushing malware to people via IDE addons, pretty good strategy for training those datasets.

Interesting - is the model that learns from (semantically indexed) code bound by its license? One could argue it is derivative work.

Interesting question. I’m not a lawyer, but I presume it’s derivative enough not to be copyrighted. For example, one could argue Google’s giant n-gram corpus is derived from many copyrighted webpages.