Live data from Hacker News

GitHub Copilot is generally available

github.blog

351–360 of 796 posts

Re: GitHub Copilot is generally available

#353
post #226

Earlier quoted context omitted.

"how to put a sentence together describing the rules, it absolutely doesn't actually understand how "Chii" melds work" The more experience I get with GPT-3 type technologies, the more I would never let them near my code. It wasn't an intent of the technology per se, but it has proved to be very good at producing superficially appealing output that can stand up not only to a quick scan, but to a moderately deep readin…

I wholeheartedly agree with your analysis, but feel like it’s ignoring the elephant in the room: writing code is not the bottleneck in need of optimization. Conceiving the solution is. Any time “saved” through Copilot and it’s ilk is immediately nullified by having to check it’s correctness. From there, the problem is worsened by the Frankensteinesque stitching together of disparate parts that you describe. I can’t i…

> I wholeheartedly agree with your analysis, but feel like it’s ignoring the elephant in the room: writing code is not the bottleneck in need of optimization. Conceiving the solution is.

I dunno about this. I know the received wisdom is that "writing the code isn't the hard part", but I think reality is more like "writing the code is only one of the hard parts". There's an awful lot of badly-written code, or code which is only partly correct, or only correct under some circumstances. The only way to make writing code not one of the hard parts is to specify 100% of the functionality, every corner case, and all test scenarios, before any code is written. And then you still have to verify that it was translated correctly into code, which I think we can all agree is another one of the hard parts!

Conceiving the solution is hard, thinking of edge cases, what-ifs, and failure scenarios is hard, creating effective tests is hard, and writing the actual code understandably and correctly is also hard!

Re: GitHub Copilot is generally available

#355
post #200

Earlier quoted context omitted.

Why should the money go the to code authors in the first place? All training data is available under permissive licenses. Assuming you're not overfitting on specific code sequences (which would require attribution - and yes, I'm aware Copilot is not immune to this problem and it needs fixing), I'd say this is fair play.

Unless something has changed, the training data also includes copyleft code, not just permissively licensed code

Regarding the training of the model - I don't think a copyright can restrict reading, and training is reading, not distributing any original data.

About deploying the model - it just needs to filter out verbatim exact snippets so it only outputs original, unattributable code. That can be done by hashing ngrams and a bloom filter. The vast majority of code generated by Codex is original anyway.

By the way, Codex is good for many other tasks, like, parsing the fields of a receipt, or extracting the summary of an email, or generating baby names, it's an all purpose NLP tool. Just call it like a function. Code completion is just one thing it does. It talks pretty great English, can compose poems.

Re: GitHub Copilot is generally available

#356

What I really want is a one-shot learning tool, which I teach once how to apply some code-transformation, and then the tool can apply it everywhere in my code.

I don’t know what language do you work with but do you mean something like a ESLint for JS/TS?

Re: GitHub Copilot is generally available

#357

Earlier quoted context omitted.

So, the problem is that it produces incorrect code…

Specifically, the problem is it produces _almost_ correct code, which is worse than incorrect code because it might fool you into trusting it.

Quite. So we agree that this code is incorrect, and thus, that we have a contradiction on our hands.

To be clear: we’re in agreement that incorrect code that passes for correct at a glance is even worse than obviously-incorrect code.

Re: GitHub Copilot is generally available

#359
So I was using copilot for a long time.

10$/Mo. Is way to much for what you get.

I mostly write js/ts code.

The suggestion feature / auto-complete feature is wonky at best and leads to bugs or just bad code in the worst case.

Even when you write comments or have a function like `addOne` and you want to add `subtractOne` it will not get it right a lot of times.

Then you have the cases were it throw 50 or more lines code at you for something very simple.

Catching errors or error handling is basically non existing.

I tried it for writing tests. It bad. It does not help at all.

I uninstalled and after some hours of work I don't really miss it.

Re: GitHub Copilot is generally available

#360

Does copilot learn from and suggest patterns in the same codebase that you're working, or does it just pull from the huge pool of projects on GH? How well does copilot help with languages like Elixir that are less common? WIth TypeScript it's been remarkable, but that's one of the most popular and surely very familiar to devs and GH, so I would expect less popular like Elixir to not perform as well. Does copilot work…

I've played with it a little bit: Copilot did pretty poorly when I tried using it with Julia- it kept suggesting Python code. I suspect it would do something similar in Elixir. I'm also a vim person who doesn't want to use VS code, but I've gotten more than enough value to get into my first IDE (with vim keybindings). A lot of tedious C++ code is getting correctly auto-generated.

It has first class Neovim support, possibly a better alternative for Vim person than any IDE.
Post reply on HN