Well, if they're profiting by serving copyright or copyleft code up for others to use after stripping the associated licensing I would think they're in a bad place. Unless their side effect is (again?) to break GPL and other open-source licenses?

I would think this a risk that commercial customers would want to avoid - getting something GPLv3 or other "infecting" licenses contaminating their product due to a developer cutting/pasting code. I also think that Microsoft would not be interested in getting a result that code was obvious and so could not be protected by patent or copyright.

Anybody using FOSS should be (vaguely) concerned about this kind of legal battle. Would hate to see Linux or GNU damaged to MS's gain.

GPL or AGPL doesn't infect anything.

You choose what license you want to use. If you violate the GPL license by including GPL software in your application then you must remove the GPL software as you do not have the right to copy it in that context.

If your application is written by exclusively copying GPL code, then obviously the entire application will be under the GPL license because you didn't write any of the code.

Getting 1-2 lines of code at a time does not "contaminate" the project, unless you get 100x that in a row and all from the same source. They are almost always generic codes, short and doing just one simple thing. What kind of GPL project is largely contained in a single line or two? It's more like using words from the same vocabulary, not like quoting paragraphs from a novel.

I think it's you who are in the "vocal minority".

Microsoft Copilot is an abuse of open source. It is perhaps the greatest theft of intellectual property in human history.

We should do everything we can to defeat it.

That said, here is my Request For Comments regarding a slightly modified BSD 2-Clause License explicitly prohibiting Copilot-style use:

https://news.ycombinator.com/item?id=33458374

We need to do everything we can to defend open source software against Microsoft and the like.

From an end user perspective, CoPilot has made my coding life dramatically better than any other tool I think I've ever used. The next closest being a good IDE which added code completion, documentation for functions, etc.

What it's functionally doing is incredibly valuable and I want to see it succeed. Figuring out how to navigate licensing effectively is an important step, but as someone who writes open source, I am happy my code is helping potentially make other people's code better too.

The problem is, that it's not only open source but also private repositories on GitHub which aren't open source and might never be.

That's not their stated position, obviously it's hard to verify but from there official stance:

"GitHub Copilot is powered by OpenAI Codex, a new AI system created by OpenAI. It has been trained on a selection of English language and source code from publicly available sources, including code in public repositories on GitHub."

A repository being public does not mean that it's open source.

I love the people that spent decades railing against intellectual property laws and how you can "steal" knowledge are suddenly Disney-level militants on protecting software copyrights

I doubt any of those would have a problem if it wasn‘t a mega corp like microsoft trying to make profits of that. Microsoft has a history of being evil.

Not dramatized, evil is the correct word to describe many pf their practices.

Completely agree, especially with the evil part.

Unfortunately you can't pick and choose what copilot is being used for.

For example, it is a wonderful tool in cutting down the time needed to generate/write tests in TDD but how do you differentiate that between large chunks of proprietary code that it may spit out.

It has some benefits, but far more downsides, and Microsoft and its people in management have a history of being very evil in the truest sense of the word.

A lot of people see that as a trigger word instead and can't recognize it for what it is. What's that say about society when a majority can't recognize evil...

Completely disagree. I think that the output that it produces is sufficiently transformative to be considered fair use, and even when people coerce it to produce code verbatim - it represents a fraction of the original code base.

It would be like if J. K. Rowling went to war over the fact that I lifted a few paragraphs from Harry Potter and used them in an entirely different book.

Also this is entirely anecdata, but at my company I'd say we have about a dozen other engineers actively using copilot and not a single one of them has an issue with it. (ideological or otherwise)

I think people opposing Copilot are overly reactive, saying they fight for open source but actively trying to stop "Opener Source". Because Copilot is that - even more open than open source. So open that even open source people are angry. It learns skills and then makes them available to everyone. The gall!

Copyright should not protect ideas, only expression. The ideas contained in copyrighted works are fair use for AI to train on, even if you don't like it.

Copyright should not protect a lot of things but that hasn't stopped courts from ruling that things are protected by copyright (when they shouldn't be).

Take a look at the Chilton's Manuals, all the publishers were bought up and now its paywalled for a subscription, consists only of steps and procedures (know-how) so you can repair your own vehicles. That shouldn't be copyrighted.

Buy a 10tb disk on Amazon for ~$300, start backing up every git repo in existence

Even with a shallow clone, I'm curious what percentage of repos this would hold.

The Stack is ~3tb: https://huggingface.co/datasets/bigcode/the-stack

If you had 10tb worth of decently written code, you could probably train a competitive model.

What a bizarre post. I haven't come to a definitive conclusion on the subject, but "how can I help a trillion-dollar organization?" is very far from my first thought on it. Is this a service folks have become dependent on?

It boggles the mind. I was also in disbelief when I saw this post.

You do have to admit that it is Microsoft (and the other vendors of similar CoPilot-like services) that are on the side of "Information wants to be free!" this time. The irony of FOSS advocates reversing course on their own principles to oppose it is deep indeed.

You can't take a person's name off their work and sell it.

People who work hard to produce open source software expect credit and acknowledgement for their work. That's why almost every license has an attribution clause.

If you don't want credit for your work, you use the MIT No Attribution License:

https://opensource.org/licenses/MIT-0

Not widely used, as you might expect.

Such a short-sighted view. The grand goal of FOSS was the abolition of copyright protection for code and and the end of proprietary software, was it not? Credit and acknowledgment doesn't even figure into it.

What is a ML model that has knowledge of huge swaths of humanity's code (eventually all of it) and can serve as an assistant to build any kind of software on demand if not a huge milestone on the road to the total software freedom envisioned by RMS and GNU? People will be able to build whatever software they want, easily and at any time, and whatever the code they finally wind up with will even feed back into making the ML model even better for the future. One could still build closed source software or harp about "credit" but why bother? The ML model would replicate anything with a modest bit of effort. It might not have been the path to end-user freedom that the early FOSS proponents imagined but, if it works, that's victory.

No, that was not the grand goal by any stretch of the imagination.

"Copyleft" protections depend on copyright protections. The reason the GNU GPL can be viral is because the license is preserved from one instance of the code to the next. If credit is not being given, then how can the GPL survive and perpetuate its virality?

RMS wanted software to "be free, like air". Free software is not gained by putting it into proprietary code without licensing it. Free software is not served by filing off the serial numbers and scattering it to the four winds. The writers of free software should be given credit and they also deserve to earn a living from writing it, if they so desire. CoPilot threatens to rob them of that livelihood by side-stepping IP protections.

> The ML model would replicate anything with a modest bit of effort.

Yeah, by stealing it.

Last time someone wrote a comment on HN on this topic he insulted a library author and said something like "you are just bitter that eventually this AI software will be able to write your opensource library based on the GitHub description alone"

Well, of course it does, it can just literally search through GitHub and produce a verbatim copy.

No, one does not, in fact, have to admit that. Your statement represents an opinion.

You mean they want to abolish intellectual property by allowing proprietary code fed through the AI to be stripped of copyright? Because that would be the end of copyright.

You need to think bigger — if the class action wins, it effectively would ban training AI classifiers on internet data.

Yes, we all want to stick it to Microsoft, but that’s not what’s at stake here.

AI research companies can't be bothered to ethically source data so everyone else should just shut up and take it? It's not like the lawsuit can stop independent researchers from using whatever data they like, but putting the fear of the judge into big AI corps is a good thing.

If this class action wins then it means we can put license restrictions on AI-training. "You may use this code to train an AI model as long as you release the weights under AIGPL, and any work substantially derived from the AI output is also released under AIGPL"

If the class action loses then this will not be possible.

Internet data? Copyright, licenses like GPL MIT etc. mean nothing if it is just all “internet data”.

Whatever happened to “information wants to be free”? There is a post on the front page right now about how it’s totally okay to post copyrighted books online.

It's almost like different folks have different opinions right?

I think “information wants to be free” is more of an observation than legal or ethical stance. Even as an ethical stance where it might be interpreted as “DRM bad” I am not sure it also means “pirating good”.

For what it’s worth, I agree with you. I understand why most of the HN crowd is against this, because Microsoft is big and evil and people want to see them suffer.

But realistically, a win here would effectively ban training AIs on the internet.

You paint us as petty and simpleminded. I have complicated feelings about Microsoft; I think they've done a lot of good and bad over my lifetime. My issues with copilot are principled. I have the exact same issue with non-Microsoft efforts to ingest copyrighted works and emit derivative works to the detriment of artists. Argue against that, not a pathetic strawman.

> ingest copyrighted works and emit derivative works

The compression ratio for Stable Diffusion is 1:25000 or more, 5B images into 5GB. I think you're overly dramatic about it. Not even a single pixel from an image. And a snippet 2-3 lines long of code, that 99.9% of the time is generic, is not worth all this scandal.

There is a distinction between learning and replicating an idea vs. the expression of it. AI is entitled to learn ideas from copyrighted works even if authors don't approve. Only expression is protected. Even API names are ok.

I think the natural balance would be to set filters that restrict replication of copyrighted code. And that's it, every kind of derivative is ok as long as it does not look overly similar to any copyrighted code. But excluding code that is generic and replicated in many sources, because that doesn't deserve to be protected and it would just stunt the model.

You could try to contribute a license-finder for copilot that would detect potential copyright violations and emit valid attribution / reproduce the licenses of copyrighted works that copilot's output is derived from.

Oh, and you'll want to detect incompatible software licenses and prevent derivative works from being created with such conflicts .

In so doing, you would be helping Microsoft to follow the law and respect my copyright. The lawsuit would probably evaporate overnight.

That's a fairly sane request, but I think people can do a bit better yet.

This is a bug, and IMO the correct solution is to fix it. Preferably people should nip it in the bud at the start of the process, rather than play whack-a-mole with the outcome.

When your model gives you back exactly the same data as was put in, that's called "overfitting". It's a kind of bug (or at very least a smell).

It means (among other things) that it might actually give back the same answer in other situations where it is wrong to do so. It could then also give a different, wrong answer in a situation where this answer was actually the correct one.

It also means that the model spent all this time "learning the answer by heart" instead of the underlying pattern. This is typically a waste of perfectly good training time if nothing else.

And of course if you have an overfit, in this case you end up suggesting copyrighted code to the user instead of something original. This is rather undesirable.

In one of the cases I looked at with a friend, we think many people had actually copied one person's code without attribution. So the training set might contain -say- 100 copies of that same solution. That would then be a strong signal that there is only one correct/common solution to that particular problem: "Copy this one guy and leave off the attribution". Oops.

I wonder how many complaints about copilot are actually +/- "blaming the messenger". I suspect that fixing/improving/mitigating this particular set of bugs may have an unforseen side effect: a number of rather interesting conversations with some of the developers in the source data-set.

The dataset is so large they only do one single pass, and the content is deduplicated. A study showed that it's usually duplicated content that is overfitted and later regurgitated. Examples that are repeated in many places might have slight variations that help them pass the filter.

Interesting!

Do you have a link(s) to the study, or more details on training?

Deduplicating Training Data Makes Language Models Better

https://arxiv.org/pdf/2107.06499

Quantifying Memorization Across Neural Language Models

https://arxiv.org/pdf/2202.07646.pdf

You don't have to do a thing. Microsoft (and Amazon, with it's CodeWhisperer, and every other similar business building such services) went into this fight with their eyes wide open. They knew this would happen and are well prepared for it. And even among the FSF's own invited whitepapers on CoPilot (https://www.fsf.org/news/publication-of-the-fsf-funded-white...) was one ("Copyright implications of the use of code repositories to train a machine learning model") that that seemed to conclude that CoPilot-like services could defensibly say that copyright was not violated regardless of the license.

So if this suit succeeded, I would be very surprised. Even the FSF isn't rushing to file suit and they care about this controversy probably more than anyone.

If you find a way, I'd like to join you.

The best thing you can do is write OSS and not care who uses it or for what purpose. The more people writing truly free and open software the better. In my mind that is the ultimate end goal of OSS.

Copilot is to Open Source what Open Source is to Closed Source.

The best thing you can do is to stay out of the way. Microsoft will hire very talented lawyers who will craft a strategy designed to maximize the odds of winning.

You have no chance to increase those odds; but it’s possible you could reduce them.

I kind of see where the OP is coming from. HNews has been dominated with posts to do with DALL-E and Stable Diffusion and expressing joy at the ease it has made generating concept art, logos, etc. The small volume of artists expressing their displeasure? Glossed over.

Now that AI is showing face at the first step of removing programmers and allowing business-speaking people to generate code suddenly there's websites and links all over the HNews front page trying to shut down Co Pilot.

Buy Microsoft shares perhaps.

You want to defend the big company? Why? Copilot is obviously breaking terms of licenses. At the very least it should only be trained on MIT and return MIT code

It does not seem at all obvious to me why training an AI on public code examples should break any license terms.

The "public code examples" have specific licenses that aren't being met.

You could buy a million copies of Microsoft Office, I think that would help.

Whether or not you care about your code being used, other people do care about theirs. They have a right to represent these concerns and see how they fare in the legal system.

At any rate, let me try to cut through your visceral anger and invite a discussion. :-)

"Is ML training sufficiently transformative that the model should be considered fair use?" - a burning question in many of our minds (maybe)

Firstly, this isn't yet clear from a legal perspective. Whether or not you agree with the class action, we need to clarify how IP law views ML training as it regards source -> model.

Anyway, here's a thought experiment: given an extremely overfitted model (to keep it simple, we can achieve a basic overfit by training only on a single piece of source data), predictions will generally replicate the training data. Indeed, one could accuse an overfitted model of being an over-engineered copying algorithm.

Therefore, the question of whether outputs are "sufficiently transformative" must at least consider the fit achieved by an individual ML implementation.

But what about a well-fitted model that generalizes to novel data? Back to the thought experiment. Let's expand and diversify the training source. In other words, we're cutting the training data with other training data to end up with a less "pure" (more general) result. Training time and the complexity of intermediate layer data (e.g. weights) skyrockets but, if you can reduce loss enough (which happens to be the goal of ML algorithms), we'll still end up with a model which reflects the aggregate training source. There's no magic: ML algorithms simply create an incredibly large/complex statistical model based on mathematically visible patterns in the source material. If the goal is to descend the loss gradient, the identity of a model is (100 - loss)% relative to the aggregate training source.

So, as for the "sufficient transformation" question, I think this question is clarified by considering ML training to be a form of "dilution" (in pursuit of model generality): if I dilute source material with other source material from different authors and compose a resulting work, am I cleared of all copyright responsibilities to the individual source works? How much dilution is necessary to put me in fair use territory? Personally, I think these sorts of cases can only be litigated on an individual basis. Also, a fairly obvious solution to clear this question is to generate your own training data, whether or not you believe this should be required.

But all that stuff isn't really what this class action case is specifically about. The root question of the case is this: do FOSS license holders reserve the right to prevent (or at least opt-out) other entities from building commercial products out of their licensed code without honoring those licenses and/or providing any compensation/attribution?

I believe we do reserve this right, and it doesn't require quashing Copilot or kneecapping the (IMO limited) potential of AI/ML. It simply requires clarifying laws and licenses accordingly and then determining whether damages have already occurred. The class action lawsuit is an assertive step to check this via the legal system, particularly considering that Microsoft is a giant corporation making gobs of cash off of this model. It's a way to answer questions and solve problems in the real world.