It's pretty wild to see tab complete go from just variables and quotes to writing entire comments and finishing functions. I like that it will learn from the project that you're currently in so it'll suggest comments and variable names that follow the standards you're using.

Edit: so my takeaway from this thread so far is that this implementation of machine learning for code completion is awful because it's trained on all the code on GitHub and will frequently create buggy results and literally steal public domain code and put it into private codebases without respecting licensing.

Having used it for a day I can say that there's some things it does that are really helpful, mainly the templating for docstrings and autocompletion of comments (e.g. # Find all minima in > {list variable} and return list of minima where everything after the > is autofilled) as well as auto filling function calls when variables you're going to pass as arguments are initialized (eg. sum_list = [1,2,3,4] total = sum( > sum_list) where > is the start of the autocorrect)

There also needs to be a limiting of the machine to only the code within the libraries in included in your codebase and the code you've written (which is already an aspect of it, where it caches something like 1kb of cute and will use it to autocomplete future cute in your project).

The creation of new code from comments is an anti-feature.