Loads of devs I’ve spoken to, from junior to principal level, absolutely love Github Copilot though. Don’t know who is paying it for them, nor if Kite was significantly worse, but I think that at least Copilot has a brilliant future ahead of it.
If Copilot generates the exact same code as the source, I don't see how that process could be exempt, it's like using the clipboard on your PC with extra steps.
1) It's the code license that matters, and the depends on what license the code the developer is building is using
2) License compatibility is undecided (see eg the different views of Apache Foundation and FSF on the compatibility of ASF 2.0 and GPL)
3) It's easy for someone to deliberately produce code that violates a license. MS is a big target and you can bet license trolls would chase it on that.
3) is not a good reason. "Big fish" is not an excuse to not provide protection. The burden to protect others from harm as a direct result of a feature of a product is on the owner of that product.
Do you mind explaining why? What do you think will happen?
This is a serious question, I'm apparently just unaware of the horror-stories that can come out of breaking copyright. (Not from the US so..)
For me its most useful for helping with bash scripts and small simple stuff, just saves a huge amount of time I would spend googling and checking small things. Not sure if copyright is relevant there or not, it certainly isn't something I am worried about. Interested to hear what your fears are based on.
ML text and image generators have a habit of regurgitating training data, especially if you happen to use certain inputs or prompts that also match said data even superficially. For example, type in "fast inverse square root"[0] and get out a bunch of Quake source code.
If you use code or art generated by an AI that was regurgitating training data, you can be sued for copyright infringement.
The way that AI gets training data these days is... questionably ethical. It's all scraped off the web, because the people who make these AIs saw court precedent for things like Google Books being fair use and assumed it would apply to data mining[1]. Problem is, that does nothing for the people actually using the AI to generate what they thought were novel code sequences or images, because fair use is not transitive.
[0] This won't work in current Copilot because a) I'm misremembering the comment phrasing and b) they explicitly banned that input from generating that output.
So, this makes sense from the perspective of writing code professionally for a company. I don't buy that the risk is significant for an individual programmer, but I suppose it is plausible that lawsuits could pose a threat to a large organization.
But I find I get most pleasure out of using CoPilot for personal tasks, everyday stuff that isn't for work. Surely there is no issue there?
And to counter the note about the scraping sources-- I have recently (out of work) been experimenting with using CoPilot to translate open-source MATLAB code into python and zig. If I make something useful in this way, it will be open source and I will credit those who wrote the original code.
Surely a use-case like that last would have to tick all the moral and legal boxes?
It doesn't block your ability to learn any more than any auto suggestion systems. But I guess it depends on what you value in terms of learning, for me, copilot allows me to focus on the larger architectural problems while not having to worry about the exact syntax of certain things (DSL query language, middleware express, typescript def annotations, etc).
Every time I don't have to context switch to look up some technical errata in my browser is a complete win for me.
I observed one guy become very resistant to reading the docs as they believed copilot obsoleted them. Programming in Ruby, the copilot suggestions will run but since you didn't read the docs you won't know that the output doesn't actually work or doesn't do what you think it does since it was intended for a different situation.
I use it. It's good for boilerplate code or basically anything where you can avoid looking at the docs. For example, I was writing an ML training loop and it correctly filled in the rest of the function after I wrote the first few lines. The code is basically what's in the pytorch docs, just fit to my model and scenario.
A coworker of mine uses it, it almost seems like an unfair advantage but I can also use it too if I want to. Specifically writing hard to decipher code eg. functional composition but Copilot will spit it out.
I haven't wanted to use it personally. I'm also not a senior dev or anything so my opinion is not worth much.
Everyone I've spoken to who has actually used it thinks it works extremely well. This includes pretty experienced developers.
I've never heard anyone claim it blocks the ability to learn - if anything it's the opposite. Many people like how it shows you APIs you weren't aware of.
I used it for about two weeks total, over two different periods. Mostly Kotlin, Java, Typescript, small amounts of Groovy. I ended up turning it off. In my experience, it has moments of utility, but most of time it felt like it was getting in the way. The kinds of things it completed well were not the kinds of things that cost a lot of time or mental energy. I found wasting more time trying to fix things that when it got close but not quite there, than it saved in spitting out boilerplate.
But I did find that I need to turn it off at certain stages when learning (or re-learning) a programming language. It's seemed counterproductive until you have a good grasp on the basic syntax of the language. But the "showing you new APIs" does seem to be a thing that actually helps.
In general, you should not be accepting code completions that you don't understand. I'm usually stricter, in that I only typically accept completions that line up with what I was planning to type anyway.
It's quite poor for learning a totally new set of programming language grammar.
If you're a Javascript developer of 5+ years of experience, it's probably absolutely fantastic. Not so great if you're learning a new programming language.
Please remember that Guido van Rossum is now employed by Microsoft. Yes, I am a fan of Guido van Rossum and his work, but I am always suspicious when famous people recommend their own company's products. It feels like excellent marketing/PR that is hard to resist.
I _love_ Copilot but the only reason I use it is because I qualify under their open source developers program, I just can't justify paying $10/month for it.
It saves a bit of time, but doesn't seem to make a difference on time to market of features, products, improvements or bug fixes.
In my experience, it's a quality of life improvement, but the things that dictate actual time to market is bottlenecked by things that aren't solved by copilot, such as overall design, decision making, requirements gathering, code structure/architecture, solution ideation, user acceptance, infrastructure setup, etc.
I think if it eventually could help with those other tasks, you'd see time to market gains, and that would start to make it really valuable.
