Ask HN: Anyone has good experience with GitHub Copilot?
I have seen some trivial examples but how far does it go. Have you been surprised by the complexity or even creativity of it's suggestions? Do you use it on a daily basis?
It’s mainly good at writing boilerplate, but a surprising amount of code is boilerplate. A lot of small algorithms in helper functions; code which is very similar but not enough where abstraction is worth it (like similar but different keywords everywhere); code which is a really straightforward translation of the above documentation or comment, and I can just have Copilot generate the code from the comment; etc.
It’s also really good at generating a “starting point” (for a function, class, etc.) which even if incorrect, helps me get over the hurdle of writing the correct version. Like, sometimes I just have no idea what to write, and after it writes some random code I realize and edit the code to do the right thing.
It’s also really good at generating code for an unfamiliar language or library, because it will use the syntax and API and even if the semantics are off, I can understand the code once I see it and correct them. Whereas if I were to write the code myself, I already know what to write but I don’t know how to write it because I don’t know the syntax, method names, and conventions. It even understands the syntax of niche languages like Coq, although obviously the completions are much worse than common languages like JavaScript.
And yes it makes careless mistakes, but I actually make careless mistakes more often (like swapping “true” and “false”, I put false into something which should obviously be true and vice versa…I’ve done that multiple times. The AI rarely makes obvious mistakes like that, instead it’s mistakes tend to be semantic errors and API hallucinations). And it can’t handle anything non-trivial (unless it’s a very well-known algorithm); and if you start typing in the wrong code, it won’t correct you, and will just generate more wrong code (it hallucinates like ChatGPT). It’s not a programmer replacement. But it’s really good at generating boilerplate, and a lot of code is boilerplate, so it’s extremely useful.
It’s actually kind of scary because I’d hate not being able to use it. There are alternatives (CodeGeeX and Codeium), but I haven’t heard anyone’s experiences or comparisons. Fortunately Copilot has been out for a long time (since 2021!), seemingly hasn’t changed much (main differences are better context and more places for code completion), and it uses GPT3 (!); so if it ever does get worse, I have more faith in an equal-performance replacement than something like ChatGPT.
It worked like a charm. Really useful even when i have to correct code. But in the end it helps me makes things about 30-40% faster.
I was surprised by the fact that if you open several tabs on visual studio code it takes into account all that extra code to give you a result.