> I also don't like Microsoft not including their own code in the model

From allegations I have read across the Internet, Microsoft might be doing those who use Copilot a favor.

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.

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.

[1] In the EU, this practice is explicitly legal

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 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.

I think copilot works very well.

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.

What about Karpathy, nobody pays him in principle.

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.