It's ML autocomplete, with help from 'traditional' autocomplete methods that use static analysis. Instead of just completing one token at a time like traditional completers, it can do entire sentences or multiple lines of code in one go, and is freakishly accurate. And since it parses language it helps you write comments, and can understand relations between code. E.g. if you are writing 2d movement code and you do x += dx it'll automatically suggest y += dy for the next line based off of previous similarities; of course if you have x += [complex math formula] it'll fix it up for y and convert cos to sin, etc.
Support for many editors, and easy to install in vim. Free for personal use. Works for all languages, including plain English (and maybe other non-code languages?).
I tried this when it first came out but it didn't seem much better than PyCharm's usual suggestions (which are admittedly excellent among its IDE peers). I rarely to (maybe never?) saw it do any multi-line suggestions, let alone accurate ones. It was also very slow to suggest anything in the first place (I believe the network calls were slow iirc, at least compared to local/native autocomplete)
Maybe its progressed and I should try it again today.
Not sure when you started, but I've been using tabnine in Pycharm for a few months and it is absolutely mindblowing. I've had long line autocompletes (no multi-line) and often time it suggests things I may not have thought of, "now that you mention it, I DO want that idiom". It's snappy enough for me and I am not exactly a patient individual.
It's only too much if you need that memory for something else though. If it manages to be responsive with that memory and you only have 1 ide open, I don't think 3gb should be a problem on a modern system.
It always starts with this premise, suddenly every application no matter how silly demands 3gb. I understand that progress often happens by putting more attention to other aspects than economics, but again some people may value that progress less.
Disappointing to see the pricing model has changed; there used to be a license for unlimited project sizes, but now appears to be a $15/mo sub for their paid service.
Understandably, this was before they transitioned from on-device models to more complex, larger cloud models.
The free 400KB limit is quite generous, but you may need to spend time tuning the ignore if you have junk in your project folder.
This thing is fucking magic.
It's ML autocomplete, with help from 'traditional' autocomplete methods that use static analysis. Instead of just completing one token at a time like traditional completers, it can do entire sentences or multiple lines of code in one go, and is freakishly accurate. And since it parses language it helps you write comments, and can understand relations between code. E.g. if you are writing 2d movement code and you do x += dx it'll automatically suggest y += dy for the next line based off of previous similarities; of course if you have x += [complex math formula] it'll fix it up for y and convert cos to sin, etc.
Support for many editors, and easy to install in vim. Free for personal use. Works for all languages, including plain English (and maybe other non-code languages?).