GitHub Copi­lot inves­ti­ga­tion

The narrative here seems to be a David and Goliath story as Microsoft profits by stomping on the defenseless open-source communities. There's two problems with this story.

First, the huge majority of open-source projects are at no real risk because Copilot offers something totally different from what they offer. Open-source projects generally take highly-complex domains and expose them as simple interfaces or executable programs. This encapsulation is where the value lies.

In contrast, Copilot just dumps code. Never once doing front-end work have I thought "if only there was a way to dump verbatim React internals directly into my codebase." In general, Copilot only replaces tasks I would have otherwise done myself.

The second problem is the biggest loser if Copilot gets shut down is not Microsoft, who can easily take the loss in stride. The real loser is the community of developers, many of them bootstrapping their own projects or trying to develop open-source in their precious off-hours, for whom every minute counts, and for whom tools like Copilot can be the difference between success and failure.

So why is he complaining about copilot and not the thousands of GitHub repositories redistributing his code with improper license? Typing the little code snippet he showed into copilot is analogous to typing it into the GH search bar and grabbing a properly-licensed result.

I don't like that opensource code is being used in a commercial product. I feel concerned about NNs learning about stuff they aren't really "supposed to" learn, because somebody published something by mistake a long time ago. But this general argument about reproducing copyrighted code is stupid, and actively trying to shut Copilot down because of that is why lawyers are cancer.

Basically, what Copilot (or anything like that) is supposed to do is to speed up your work, i.e., ideally, to write exactly what you'd write, but orders of magnitude faster. How do you write code? Well, you may have a solution in mind — if it's something really original, rest assured, Copilot won't guess it. It can only hope to guess something that, in a sense "has a correct answer" to it. In fact, it does it worse, than it should be: graph traversals, matrix operations and such should be guessed flawlessly (in a perfect world every PL would have some primitives implementing them in the best possible way, but ours is not perfect). If you don't know how to traverse a graph, you'll go and look for a reference. 15 years ago it was likely a book, then looking up on the Wikipedia or StackOverflow became way more likely. For the last 5 or so years literally searching it on GitHub became viable because of better search engines and the sheer size of it.

Now, if I found a matrix transpose function in an open-source project, which I cannot include as a library for some (usually technical, but maybe not) reason, so I memorize it, close the page and re-type it in my IDE, do I have to be restricted by its license? Then, doing so is obviously stupid, so how about me just copy-pasting it, while renaming some variables so that the teacher wouldn't notice? And, given that this is not my homework, there's no teacher and variables are named perfectly as they are — doing that is also really stupid, so I might have just copy-pasted it. So, how about now, do I have to publish my code under GPL3 now? Is this theft? If any lawyers say yes — fuck these lawyers. It is nonsense.

https://arxiv.org/abs/2106.06981 - Thinking like Transformes

The paper suggests that transformer work on a set of select, aggregate and element-wise operations. Which seems pretty close to the SQL statements i write from day to day.

> It's very obvious from enormous numbers of examples that current AI systems are capable of learning much more abstract features of human culture

This doesn't seem to translate well to code. You can't copyright 'rembrandt's style', which is what dall-e and co learn from analysing those paintings. But what copilot gives you is a sizable chunk of code. That code is (mostly) exactingly precise: It's not like the AI learned a style and recreates the style. It learns what you intended to do and then verbatim copies a code chunk in. I'm pretty sure the AI part comes in to determine what it is you were likely trying to do, so that it knows which code to copy. Not to generate the code out of the AI model. At least, that's how I understand it works.

That is the fundamental difference.

> OK, if you said "what word comes next? MR. AND MRS. DURSLEY OF NUMBER FOUR PRIVET DRIVE WERE PROUD TO SAY" ... same thing

Let the AI fill in the rest of that sentence and we can debate whether that is copyright infringement or not. However, if the AI system is capable of finishing that sentence, it can presumably fairly trivially be asked a slightly different question. Instead of 'finish the sentence', how about: "Suggest the next likely sentence". That system would presumably generate the next exact sentence straight from the book, and keep going and - voila you recreated the entire book.

Which is clearly copyright infringement.

AI systems have a 'volume dial' to configure how much they mix and match. Turn it down low enough and asking DALL-E for 'a girl with a blue bandana and an earring in the style of Vermeer' will just give you a copy of Girl with the Pearl Earring, reproduced sufficiently accurately that it's trivially a copyright infringement (let's leave out the notion that the painting has moved past its copyright date of course).

Point is, for copilot, the volume dial has to be kept extremely low, because you can't just mash 5000 snippets together, unless those snippets are identical. Which is its own intriguing copyright infringement conundrum (500 artists each individually paint the same thing, and they can each prove they weren't influenced by the others, thus, not copyright infringement. Then, you reproduce the averaging of the 500, which results in yet another painting. Did you just infringe copyright? Surely the answer is 'yes', but whose copyright did you infringe? All of them? Is it 'yes'?)