I mean, the Tab completion in Cursor is already seriously good. Sometimes it’s almost as fast as an agent, but way more reliable. If you’ve got a super clear idea, you can just write a tiny bit of starter code, hit Tab, and it finishes the job. It’s fast and super accurate, without the unpredictable weirdness you sometimes get with agents.
So, what’s the next big thing for Tab? As a developer who uses it every single day, I’ve got a couple of thoughts I wanted to throw out there:
-
Context Size is Still a Thing: It seems like Tab’s context is still a bit limited. If I haven’t opened a file in the editor, it can’t correctly predict code that needs to reference it. To get those high-accuracy predictions, you basically have to open up the relevant files first. It looks like Tab is mostly using the code you currently have open as its context.
-
Project-Specific Fine-Tuning: What if we could train or fine-tune on top of the general Tab model specifically on our own projects? That would create a custom Tab model just for us, letting it understand our codebase and needs way better, leading to much more accurate suggestions. Technically, it should be possible, and honestly, if that feature existed, I’d definitely be willing to pay for it.
What else? I’m kind of drawing a blank right now because Tab’s prediction speed is already crazy fast. What are your thoughts? Anything else you guys are thinking about?