> [we] use copilot basically a line or three at a time as a smart autocomplete.
I think this is the best way to think of CoPilot. GitHub is selling it like its going to write all your code for you, but in reality it is just next-generation auto-complete.
That's not a bad thing. In some ways I'd argue its actually better. GitHub needs to change its marketing because even most developers seem to think that its out there to take away our jobs. Its not and can not. But it provides the smartest auto-complete you've ever seen and that can be useful, especially when wading through mundane parts of your codebase.
I found copilot to be less useful than autocomplete. Typical autocomplete suggests things that actually exist and work in the codebase. While Copilot would suggest things that look superficially like names that exist or ways I might have named it, but very often just wrong.
I find typed languages like Rust or Typescript make VS Code super powered and provides much more value than copilot.
Not surprisingly (Google's internal code analysis tooling is quite sophisticated), Google has added a post-processing filter to remove results that don't compile, but that's not publicly available: https://ai.googleblog.com/2022/07/ml-enhanced-code-completio...
I think this is the best way to think of CoPilot. GitHub is selling it like its going to write all your code for you, but in reality it is just next-generation auto-complete.
That's not a bad thing. In some ways I'd argue its actually better. GitHub needs to change its marketing because even most developers seem to think that its out there to take away our jobs. Its not and can not. But it provides the smartest auto-complete you've ever seen and that can be useful, especially when wading through mundane parts of your codebase.