Hacker News Reader

Comment by throwaway675309 | Hacker News Reader

Even if this succeeds (I sincerely hope that it does not, as I believe these to be sufficiently transformative as to be considered fair use), it doesn't functionally matter in the long term:

1. The ability to be able to both run and train these models is going to eventually be plausible on a home machine with even modest hardware.

2. These models, e.g. such as one scraped from all of the code on GitHub, will be publicly available via torrent or whatever.

3. People will be able to run it locally as an integrated plug-in in their IDE of choice, jetbrains, vs code, etc.

4. You'll never know if somebody has lifted a bit of code in violation of a license anymore than you would be able to tell if somebody copy-pasted from stack overflow without attribution in any commercial application

View on HN

It's not transformative at all, nor quoted for study/discussion, and that's exactly what makes it a problem.

If the snippets merely had attribution, they'd be both legal and acceptable to most open source authors.

The code is not transformative because the quoted code is not used for some other purpose like as part of an article discussing whatever the code does, it is used to do exactly it's original job.

If you wrote a book by quoting choice paragraphs from other books, without crediting any of them, presenting the new book as simply your pile of deep insights, that is not transformative, even though your book is not the same as any single of the source books. It's also not fair use, even though all the quoted bits are short, both because of the usage and the lack of accreditation.

We could have something like copilot just fine if done above board, but as it is right now github are simply outlaw and with no excuse at all.

I read these kinds of comments on HN and I wonder if the people complaining about copilot have actually used the thing for more than 5 minutes?

Most of the time copilot is giving you one-line completions that are heavily adapted to the surrounding code. It’s not “writing a book by quoting choice paragraphs from other books”. It’s fancy autocomplete.

You can direct copilot to give you anything you ask it for. If you ask it for an entire function, and you don’t give it any context, then you are more likely to get a snippet from inside its memory. But even when this happens you are unlikely to keep the entire snippet unmodified because the chance that it is what you were looking for is low.

> you are unlikely to keep the entire snippet unmodified

Does that matter? It's not the end users who are being accused of copyright violation, it's the tool itself.

If the argument is that the tool is facilitating copyright infringement by users then it matters whether or not the users actually keep the supposedly infringing code, yes. If the code gets deleted immediately afterwards by the user then the claim of infringement by the user is farcical.

If the argument is that copilot is infringing because it merely displays an allegedly copyrighted snippet for a few moments, even if the user does not accept the completion, or accepts it but then deletes it, then that is a completely different argument because there is very little identifiable damage to the authors, which bolsters an argument of fair use.

The tool itself is infringing copyright. It possesses and reproduces copyrighted code without it's attribution. Full stop.

If I steal your television, it doesn't matter how much I get for it at the pawn shop.

You can Google and find GPL'ed code. That doesn't mean Google is facilitating license violation even though the Google search index posses and reproduces GPL'ed code.

> It possesses and reproduces copyrighted code without it's attribution.

To be clear - attribution does not fix GPL license violation.

You can Google and find GPL'ed code.

Emphasis mine.

This is different from asking Google and getting code directly from Google without the license - that is, without any indication that this code was taken from somewhere that has a license on all parts of it, not just the parts that Google didn't bother to show you.

> it matters whether or not the users actually keep the supposedly infringing code

That's not how this works.

If I give you a blob of uniformly random bits and then give you a passphrase you can use with a tool to turn that into a Hollywood movie, and you do so, and then you delete it, that isn't not copyright infringement.

> If I give you a ... Hollywood movie, ... that isn't not copyright infringement.

It is.

I'm no lawyer, but judging how people reacted to DeCSS (09F9 1102 9D74 E35B D841 56C5 6356 88C0, anyone?)… why do you say this isn't copyright infringement?

Copyright doesn’t go away simply because you edit what was copied, deleting it may work assuming you don’t recreate to much of the original work.

There is a concept in copyright law of derivative work which means simply editing Harry Potter isn’t enough to publish it without paying for the rights to do so. Even highly transformative fanficfion can run into this issue when much of the original work’s characters and setting remain.

Disney gets away with this stuff by copying public domain works, copilot and it’s users don’t have that defense.

The key word is fan fiction. Almost every single word in fiction is a creative expression and thus copyrightable.

But works of non-fictional literature (of which source code is an example) have to work harder to justify their copyright.

Many of the snippets of code generated by copilot contain zero creative expression. They are just mechanical implementations of an algorithm.

You mentioned setting and plot. The source code equivalent is architecture. Copilot generated snippets are usually too short and too algorithmic to have their own architecture.

What makes a derivative work derivative is that it copies creative expressions from the parent work. If you edit them out, then it’s not a derivative work anymore.

But what if not has the algorithm be designed for to extract that creative expression from the original work.

The use is only helping developing the algorithm by adding more value to the product in improving extracting the creative expression.

Copilot is a loose/loose game for developers, regardless on which end they are.

Variable names are just as subject to copyright as character names, and specific implications of algorithms can vary quite a bit. Just look at the homework assignments of any introductory programming course.

Which is why even very short snippets of code have been an issue.

Pretty much as if you buy a bike and it turns out it was stolen it's still not your bike someone who blindly accepts what turns out to be a chunk of someone else's code can still be unknowingly guilty of copyright violation.

>Most of the time copilot is giving you one-line completions that are heavily adapted to the surrounding code.

Keyword being "Most of the time"

There are countless examples of people just letting Copilot's "autocomplete" run wild and it ends up reproducing many lines of code, often verbatim, and not just single lines of code.

A lot of people seem to think changing the variable names makes it new code.

The count is like 5

Considering how easy it is to generate additional examples, I think it's higher than 5. Five big, popular blog posts, maybe; but there are way more than 5 examples just on Twitter.

You need to bait Copilot really hard to make it output that. To the point where it would be much easier to locate and copy-paste the code directly from the source yourself. It doesn't happen accidentally and it's not how you normally use Copilot.

Most of the time I enter a bank I don’t rob it

Sure. I used it for about twenty minutes before I turned it off because it seemed clear to me that it wasn't much different than copying code from open source repositories.

That is absolutely not my experience at all and wish people would stop with the hysterical hyperbole. The tool typically produces code that would only make sense in my codebase because it uses internal types which only make sense in my code base. That is not and cannot be copyright infringement.

100% agree. This matches my actual usage of github copilot. All the drama around it seems to just be mostly for headlines to fuel the outrage machine.

Sure, it's clever enough to swap identifiers correctly, but the shape of the code and techniques it uses were familiar to me from elsewhere, at times.

It's not unlike copying code from stack overflow and swapping the identifiers around.

> the shape of the code and techniques it uses were familiar to me from elsewhere.

Isn't this exactly what you'd expect an AI coder to do?

That's precisely the problem with it. AI, as it is, can be considered a lossy compression technique with multiple document recovery, convolution and interpolation ability.

Can't humans be considered a lossy compression technique with multiple document recovery, convolution and interpolation ability?

What is the specific difference?

How did you learn the shape of the code and techniques that you recognized? By looking at other code? Are you violating copyright when you are writing similar code yourself?

> By looking at other code? Are you violating copyright when you are writing similar code yourself?

In that order likely a violation, but it depends on the concrete case, always.

From decades of looking at interesting code.

I take care to attempt to not duplicate code; it's why don't use copilot.

> By looking at other code? Are you violating copyright when you are writing similar code yourself?

It's certainly possible. Back when Phoenix reverse-engineered the IBM BIOS using published sources with a restrictive license, they did it by having one team read the sources and write a very detailed specification of everything that happens there, and then another team used that spec to write new code from scratch. They did it that way because if the first team were to write the code themselves, it is quite likely that the result would be legally considered derived work.

The authors of these comments have never used it. Their understanding of the issue comes from having seen a tweet where someone coaxed CoPilot into almost-reproducing a piece of code that thousands of people have copy and pasted into their GitHub repos over the years, such as the quake square root code or that matrix multiplication code.

They also usually have a Luddite axe to grind.

If they succeed in getting CoPilot outlawed (obviously they won’t), then I will rent $50 worth of GPU time and train my own, like some kind of cyberpunk outlaw.

To attribution I would add license compliance.

I agree but I'm saying that attribution would be all that's missing from compliance in most cases.

But not, I'll hazard a guess (ianal), if its a Gnu GPL license. Right?

I didn't want to make an unreadably long comment by trying to solve every detail.

But GPL just adds that the source be made available to any recipient of code you used. That's pretty easy to arrange because even a link satisfies it.

If copilot can be made to spit out attribution, then it can spit out links at the same time.

Another idea is copilot could be changed to only include code where the authors opted in to an aggregate credit where your new program only has to declare that it used copilot and a link to copilot's training set along with your programs source, without trying to itemize each bit of output.

There could also be other versions of copilot that includes other code under other terms, like pure MIT or PD code where the original author already explicitly granted usage with no terms, or paid commercial code where github paid the authors to be able to include code to be used in this way and maybe with terms where the end user does not have to re-share.

The "link" necessary here would be a link to the full source code of the software developed with the use of copilot, distributed with every copy of said software

This isn't my experience with Co-pilot's suggestions. I've literally been able to have Co-pilot suggest a complete unit test based on a novel structure I hand-coded myself and a few words describing the unit test. The constants are often wrong, but it saves minutes of fidgeting with the syntax for unit tests and assertions.

These are not quotations from other people's code but something about the deep structures of language and programming language semantics. However, I suspect if you knew enough of a snippet from other source you could coax Co-pilot to suggest code learned from that source, but it would likely be washed over by other code in the corpus where it coincided with meanings.

Worth noting with models like copilot is that if you deliberately give it an input similar to the training contents, odds are it'll near verbatim reiterate it.

The main issue is that while you can use copilot to create "new"/transformative code, it's also trivial to get it to pump out licensed works in a form where you could claim "I didn't know it was taken from x project with y license because the tool made it for me".

I personally have no problem with copilot in concept however to do it (or any other AI model based text/graphics tool) without infringing on people's copyrights is practically an unsolved problem (excluding just per-licensing the training data ahead of time).

I mean, you can prompt me (or any other engineer) to spit out copyrighted code. FizzBuzz comes to mind… as do a number of algorithms I’ve written in the past which belongs to my past employers…

I really think we are entering some interesting territory that will likely be an interesting can of worms.

There is somehow something different if you knew the entire code to quickbooks from memory, and had an api where I could request any 10-line chunk of it I wanted, as many times as I wanted.

So you’re saying that how fast someone can type and how well they can recall makes a difference? I don’t type over 120 words per min like my grandma, but I have a photographic memory. I can tell you what file & line a chunk of code belongs to, or spit it out verbatim, customized to the current problem I’m working on.

So, you’re saying I can’t work in this industry? That seems a bit harsh.

> So, you’re saying I can’t work in this industry? That seems a bit harsh.

If you're going to be spitting copyrighted code out in violation of any licenses it might be made available under… yes, you can't. Most employers would not appreciate that behaviour. But I doubt you actually do this, even though you're capable of it. You reason about your code; you're not just being a predictive text engine.

It sounds like for you in particular, yes, since you seem to want to go out of your way to find any way to violate copyright, even when the terms are intentionally generous. Indeed such a person should not work in this industry, though I'm sure there are plenty of employers who are happy to have you steal for them, so you will be able to regardless.

My point was that we all do this /not on purpose/ (and for me being an exception, can make sure I don’t personally). But when I see code that existed in another company with some variables changed, I don’t flag it. There are only so many ways to describe a chair, are they all copyrighted?

"My point was that we all do this /not on purpose/"

I don't concede the equivalence.

It's really simple. If you are outputting licensed code and not abiding by the terms of the license, then yes, that is a problem.

Companies already pay a lot of money for datasets to train models on in other spaces outside of software development. On top of that, they spend a lot of money on labelling and what not.

Software is unique in that there is a cultural trend to share source code, so that makes it easy to compile into "free" datasets.

I wouldn't say it's an unsolved problem, it's just that there are no incentives to compile or pay for datasets when Microsoft already has petabyes of code to train on. If anything, I expect Microsoft to sell datasets based on GitHub repositories if Copilot-like models survive this lawsuit and are conmoditized.

Not totally unique in that respect, the situation doesn't seem too dissimilar from the one that led shutterstock to launch their contributor fund.

Commoditized*

> The code is not transformative because the quoted code is not used for some other purpose like as part of an article discussing whatever the code does, it is used to do exactly it's original job.

Hmm..

> The printing press is not transformative because the printed text is not used for some other purpose, it is used to do exactly it's original job.

See the error in your logic? The potentially transformative part is not the code itself. It's the impact to the process of creating the code.

For copyright purposes the printing press is not transformative. No contradiction here.

Well lucky for us paragraphs of prose are artistic expression and the snippets of code returned by Copilot are utilitarian.

The only things covered by copyright are creative choices. In software, as established by Whelan v. Jaslow [0] this is the structure, sequence and organization [1].

The district court ruling in the Whelan case drew on the established doctrine that even when the component parts of a work cannot be copyrightable, the structure and organization of a work may be. The court also drew support from the 1985 SAS Inst. Inc. v. S&H Computer Sys. Inc. in which it had been found that copyright protected organizational and structural details, not just specific lines of code. Structure, sequence and organization (SSO) in this case was defined as "the manner in which the program operates, controls and regulates the computer in receiving, assembling, calculating, retaining, correlating, and producing useful information." SSO refers to non-literal elements of computer programs that include "data input formats, file structures, design, organization and flow of the code, screen outputs or user interfaces, and the flow and sequencing of the screens."

However:

The Whelan decision initiated a period of excessively tight protection, suppressing innovation, since almost everything other than the broad purpose of a software work would be protected. The only exception was where the functionality could only be achieved in a very small number of ways. In these cases there could be no protection due to the merger doctrine, which applies when the expression and the idea are inextricably merged. Later the same year, in Broderbund v. Unison the court cited Whelan when finding that the overall structure, sequencing, and arrangement of screens, or the "total concept and feel", could be protected by copyright.

For a brief overview of the merger doctrine:

https://www.lucysnyder.com/index.php/copyright-law-merger-do...

So this excessively tight protection was rectified in CAI Inc v Altai [2], which established the legal doctrine of Abstraction-Filtration-Comparison [3] which says "As we have already noted, a computer program's ultimate function or purpose is the composite result of interacting subroutines. Since each Subroutine is itself a program, and thus, may be said to have its own 'idea,' Whelan's general formulation that a program's overall Purpose equates with the program's idea is descriptively inadequate."

[0] https://en.wikipedia.org/wiki/Whelan_v._Jaslow

[1] https://en.wikipedia.org/wiki/Structure,_sequence_and_organi...

[2] https://en.wikipedia.org/wiki/Computer_Associates_Internatio....

[3] https://en.wikipedia.org/wiki/Abstraction-Filtration-Compari...

> If you wrote a book by quoting choice paragraphs from other books, without crediting any of them, presenting the new book as simply your pile of deep insights

This is quite literally how most of the Old and New Testaments came into existence. The authors of Matthew and Luke certainly didn't bother mentioning that they copied a bunch of stuff from Mark and whatever the Q source is (my bet is on either the Gospel of Thomas or some source thereof). Nor did the authors of Leviticus and such mention how they ripped large swaths of what's now known as Mosaic Law straight from Hittite laws and the Code of Hammurabi. These works are nonetheless considered pretty transformative.

> This is quite literally how most of the Old and New Testaments came into existence.

That happened a few (thousand) weeks before copyright law was invented.

> The authors of Matthew and Luke certainly didn't bother mentioning that they copied a bunch of stuff from Mark and whatever the Q source is (my bet is on either the Gospel of Thomas or some source thereof). Nor did the authors of Leviticus and such mention how they ripped large swaths of what's now known as Mosaic Law straight from Hittite laws and the Code of Hammurabi.

The Q source (two-source hypothesis and four-source hypothesis both) is speculative, and there are several reasons to criticise the theory – such as lack of corroborating historical evidence that the Q source ever existed as a separate document.

That aside, Leviticus likely got its Hittite and Hammurabi influences via Mosaic Law, not (primarily) the other way around. (Legal systems influence other legal systems? Who knew‽) Merely saying the same thing isn't protected by modern copyright; facts are not copyrightable. (There is such a thing as "database rights", but that doesn't really come into play, here.)

> These works are nonetheless considered pretty transformative.

The merits of our current copyright regime aren't terribly relevant, here. If you want to change the law, change the law; don't try to have the existing law interpreted differently in select cases, especially if you're only doing that where it benefits already-powerful entities like Disney or Microsoft.

> don't try to have the existing law interpreted differently in select cases, especially if you're only doing that where it benefits already-powerful entities like Disney or Microsoft.

Who says it only benefits the already-powerful entities, or that it even benefits them at all in the long run? As evident from my other comments on this topic, my hope (and indeed expectation) is that they'll shoot themselves in the foot and be hoisted by their own copyright-infringing petard. It wasn't that long ago when copyright enforcement wasn't a thing, and that lack of enforcement didn't seem to stop people from creating all sorts of artistic and literary and musical works, great in quantity and quality alike. The more the megacorporations backtrack on their copyright absolutism and admit that ignoring copyright expedites creative work, the more ammunition the rest of us have against their continued insistence on copyright absolutism.

Put simply: Copilot sets a precedent that I suspect will lead to the downfall of intellectual property law entirely. Probably not alone, but very likely one of several dominoes toppling toward that end result.

> and that lack of enforcement didn't seem to stop people from creating all sorts of artistic and literary and musical works, great in quantity and quality alike.

It's why several of Shakespeare's plays are missing, and others are only available in modified form. Since there was no exclusive right of performance, playwrights (and other similar creatives) had to keep their works close to their chest. Copyright does exist for a legitimate reason.

The biggest problems with copyright today, are:

• it's far, far, far too long; and

• corporate gatekeeping allows nigh-monopolistic corporations to take authors' copyrights from them, meaning copyright law doesn't protect authors at all!

(Rebecca Giblin and Cory Doctorow discuss the latter in their book Chokepoint Capitalism, which I haven't read yet.)

> The more the megacorporations backtrack on their copyright absolutism and admit that ignoring copyright expedites creative work, the more ammunition the rest of us have against their continued insistence on copyright absolutism.

I'd love it if it worked that way. I hope you're right – but I believe you're wrong. We can let them do this backtracking without letting copyright law tilt even further in their favour.

> It's why several of Shakespeare's plays are missing, and others are only available in modified form.

If copyright enforcement was a thing in Shakespeare's time, then a lot more of his plays would be missing today; there would've been far fewer "modified forms" even partially preserving works for which the original was lost, and far less opportunity to produce copies of unmodified forms.

> I'd love it if it worked that way. I hope you're right – but I believe you're wrong. We can let them do this backtracking without letting copyright law tilt even further in their favour.

I guess what I'm getting at is that this doesn't tilt the law in their favor. Either Copilot ignoring the licenses of source material is deemed legal (at which point people are free to launder proprietary code through ML models, destroying copyright law as applied to software) or it's not (at which point a multi-billion-dollar corporation loses a revenue source). Lose-lose for corporations, win-win for the rest of us.

> If copyright enforcement was a thing in Shakespeare's time, then a lot more of his plays would be missing today; there would've been far fewer "modified forms" even partially preserving works for which the original was lost, and far less opportunity to produce copies of unmodified forms.

No. His plays were preserved and published by his friends, after his death, so that they weren't lost. Because the curtains had closed, the players had moved on, and the loss of his plays was deemed worse than the loss of monopoly over them. Copyright was originally created to remove the incentives to keep things private and secret; if you have exclusive rights to copy it, and that's protected by law, you're able to release those copies without fear of somebody taking what you've made and pushing you out.

Look at the playwrights who didn't have friends like Shakespeare's. Where are their works, now?

That being said, the world has changed since the 1600s. Copyright (with reasonable term lengths!) was necessary before we had the 'net, but perhaps not so much now.

> Either Copilot ignoring the licenses of source material is deemed legal (at which point people are free to launder proprietary code through ML models, destroying copyright law as applied to software)

The law doesn't work logically like that. If Copilot is deemed lawful, there will be justification of that decision that doesn't allow us to launder proprietary code through ML models. Because copyright isn't about author's rights, and hasn't been since… the 60s? Since some time after the Universal Copyright Convention of 1952, anyway; I'm no historian.

> No. His plays were preserved and published by his friends, after his death, so that they weren't lost.

You literally just said that there are plays of his which only survive in modified forms - i.e. ones which his friends were unable to preserve. Those modified forms are the ones that would disappear in this supposed alternate universe wherein Elizabethan England had copyright enforcement - because those modified forms would've been DCMA'd into nonexistence, and thus those plays would've been lost entirely.

> Look at the playwrights who didn't have friends like Shakespeare's. Where are their works, now?

Christopher Marlowe? https://en.wikisource.org/wiki/Author:Christopher_Marlowe

Ben Jonson? https://en.wikisource.org/wiki/Author:Ben_Jonson

Francis Beaumont? https://en.wikisource.org/wiki/Author:Francis_Beaumont

John Fletcher? https://en.wikisource.org/wiki/Author:John_Fletcher

Thomas Middleton? https://en.wikisource.org/wiki/Author:Thomas_Middleton

John Lyly? https://en.wikisource.org/wiki/Author:John_Lyly

The answer to your question seems to be: they were preserved just fine and dandy.

> If Copilot is deemed lawful, there will be justification of that decision that doesn't allow us to launder proprietary code through ML models.

A license is a license. If it's okay to ignore the terms of the GPL, then it's okay to ignore the terms of any other software license, including Microsoft's EULAs. The law works on precedents, and Copilot's legality sets a precedent that threatens the very concept of software copyright.

Even if Microsoft somehow avoids being immediately eaten alive for lunch by the myriad competitors who'd love nothing more than for Windows to lose its copyright protections (they do have plenty of money to spend on lawyers, after all), their best case scenario of "we somehow managed to carve out an exception for our EULA being enforceable that somehow doesn't apply to other software licenses" would demonstrate rather plainly the fundamental inconsistency in intellectual property law, making it all the easier to justify abolishing it entirely.

Thank you for the solid arguments.

I think your opinion is bad while also misrepresenting machine learning.

> It's not transformative at all, nor quoted for study/discussion, and that's exactly what makes it a problem.

This model works in the same way as GPT-3. It's just predicting what the next most likely word will be given the previous words. It does this by creating a generalizaiton over the content that was used to train it. This generalization should be similar (or close to the same) to if it had a completely different set of training data.

In the same way that you could create 60,000 new training images comparable to the MNIST dataset, and get a model that works at an equivalent level, you could do the same with source code. It doesn't matter where the original data is from because the end result should be very similar when models reach higher levels of accuracy.

> but as it is right now github are simply outlaw and with no excuse at all.

This is an insane view on machine learning. When you read content, neural weights are adjusted in your brain to keep it as a memory. Is that copyright infringement too? Do you now owe Disney royalties because you stored their movie in your memories?

It's also an incredibly harmful view. Machine learning is dependent on data. Services like search, translation, recommendations, etc. are all dependent on training data. And, you're making the case, that since you don't like that it was trained on source code, it should all be illegal.

Incrediblest full of all the harmfulsts! Not only Machine learning:

"Watching ditigal Walt Disney Movies is dependent on data."

It is illegal as the attribution is missing. Also copyright applies etc. see the lawsuit. Its less a technical thing, more a legal one.

> Incrediblest full of all the harmfulsts! Not only Machine learning:

I do not agree that it's a good idea to make using copyrighted data for training illegal. The people who are upset that GPL code was used for training should be happy that a large corporation believes that training on copyrighted works is fair use. It's a more democratized position.

> It is illegal as the attribution is missing. Also copyright applies etc. see the lawsuit. Its less a technical thing, more a legal one.

I think in reality, this is just a lawyer or group of lawyers who realized they had a case and a reactionary audience to possibly make them millions of dollars.

You are assuming that laws/rights operate on basic properties like learning or remembering, but they operate on goals (e.g. humans can run, but there is no speed limit for a pedestrian. But if we learned to run at 100mph somehow, a limit would be introduced). The goal is to motivate creators to create on terms that would be fair enough and useful for a society. If you “generalize” their work at such scale, it is a very different situation, like when you are using a sidewalk for robots running at 100mph and demotivate anyone from using it.

Services like search, translation, recommendations, etc. are all dependent on training data

There is an obvious benefit in being represented in search results and recommendations.

> If you “generalize” their work at such scale, it is a very different situation, like when you are using a sidewalk for robots running at 100mph and demotivate anyone from using it.

I don't understand what you're trying to state.

If you make fun of someone on national TV, someone doesn't like it either. It doesn't mean that it should be illegal.

The point of attribution and copyright is to create a creator/inventor-friendly environment. The copyright law used realistic points in a problem space to provide it. These new AIs enlarged the problem space, but it doesn’t mean that the initial idea of copyright is not applicable to them by default, or should not be. It’s okay to vote for or against that, but the potential systemic effect of that vote should also be kept in mind.

I hope this makes more clear why I used this analogy.

> The point of attribution and copyright is to create a creator/inventor-friendly environment.

They're building a class, but, beyond that, I'm not really confident there are any parties that feel like they've been harmed by what they're claiming is a violation. Microsoft didn't just copy repositories from GPL licensed code, but from other companies too. Why didn't these companies care?

It is a good counter point though, as Waymo has released datasets that can only be used under a license. While at the same time, GitHub did not initially use discretion for people who had agreed to their terms of service and were using their hosted repositories.

While I think the legal system might lean towards enforcing consent to use the data in certain circumstances, on the technical side, if it's really a generalization that is resulted from training, the initial data is meaningless in the final model. It could be argued that it did not harm this company economically because the model trained without the data would still have an almost equivalent financial impact.

> It’s okay to vote for or against that, but the potential systemic effect of that vote should also be kept in mind.

My position above was clear, I do not think the law should be changed yet. The systematic change is from someone stating a violation happening when I do not believe that it is a violation. This lawsuit could result in models like Stable Diffusion becoming illegal, which I view as incredibly harmful to the future of artificial intelligence research.

What if we extrapolate a little into the furure. Copilot-likes and SDs become useful tools for knowledge/mastery extraction and application, much better than today. Will people want to create new knowledge, or will they be satisfied with reshuffling existing one? Will opinions change if people realize that there is no point in creating something new (it will be generalized), and existing is easily reproduced without them (classic “took our jobs”). Can it all happen? How do you see it?

> This model works in the same way as GPT-3. It's just predicting what the next most likely word will be given the previous words. It does this by creating a generalizaiton over the content that was used to train it. This generalization should be similar (or close to the same) to if it had a completely different set of training data.

It's not enough of a generalization to be acceptable IMO.

Let's see GPT-3 as something analogous to a brain. Human brains can do multiple things:

- learn through experience and re-use concepts, but writing completely original code,

- copy a program's structure without copy-pasting the code, effectively re-writing it,

- copy the exact code, with a few minor details changed.

Currently, Copilot does a lot of the last point [1]. So just like humans can choose different actions and the legality depends on the action, ML models can do different things, and these aren't automatically legal by virtue of coming from an ML model.

It's not enough to argue about how machine learning works. ML can do tons of things, and GitHub's current methodology leads to something close to copy-paste. Maybe it can learn a semi-original way to write common things like searching in a list, but for more exotic uses like complex algorithms, being unable to actually understand how to code, it basically has to act as a search engine for existing implementations.

> Do you now owe Disney royalties because you stored their movie in your memories?

That argument makes no sense. Nobody would be complaining if GitHub had just trained a model on that code. The potentially illegal part is offering a service that creates "new" Disney-like movies by assembling parts of Disney's IP.

[1] https://twitter.com/docsparse/status/1581461734665367554

> It's not enough to argue about how machine learning works. ML can do tons of things, and GitHub's current methodology leads to something close to copy-paste. Maybe it can learn a semi-original way to write common things like searching in a list, but for more exotic uses like complex algorithms, being unable to actually understand how to code, it basically has to act as a search engine for existing implementations.

I don't disagree that it is over-trained on certain sequences of words, but, overall, I think the generalization fine. It's often pre-prompted with content it has not seen before, resulting in unique new content. There's nothing copy-pasted about this, just a statistical understanding of what usually happens next.

If the pre-prompt is something very specific, ex. "a dog runs in the park while during the middle of the day, while the sky is the color _____" it will obviously output blue. The same can be true when there are only very specific known algorithms that have been used more frequently than others. And, in the cases where it does commit something arguably comparable to copyright infringement, it would probably on the programmer and not on the model for deciding to use it.

> The potentially illegal part is offering a service that creates "new" Disney-like movies by assembling parts of Disney's IP.

In very specific cases only does it output copyrighted content. Most of the time it is just outputting a generalization of what is expected. It isn't Disney's content, but human content. Also, just creating content that has some similarity isn't copyright infringement. Satire has been well accepted as fair use.

If you wrote a book by clipping small bits of others' work and assembling it into something cohesive that would most definitely be fair use.

Incorrect, if you did what I actually said, which mimics what copilot does.

> You'll never know if somebody has lifted a bit of code in violation of a license anymore than you would be able to tell if somebody copy-pasted from stack overflow

Sure, there’s a sense in which copilot is just an alternative way of finding these snippets; if someone used Google to find a piece of code then chose to copy-paste it, the result would be the same. The code is out there on the internet, being indexed - and incorporated into training models.

The difference is that Google will (probably, if they don’t just yank the content into a snippet on the Serp) link you to the GitHub source, where the license should be clearly visible, so if you copy paste it you know what you’re doing. Copilot produces the code without attribution.

But it’s a narrower difference than we might actually wish. What is it about google’s pointing you to a line of code in a GitHub repo that absolves Google from any responsibility to tell you that the code it sent you to might be copyrighted? Presumably, when you land on that page you are expected to recognize from context that this is probably licensed code, and go look for the license terms. Why can’t copilot also rely on that same ‘user must recognize that some of the results produced might be copyrighted’ logic?

Worth noting that Google image search has long presented its results under a caption pointing out that the results may be copyrighted and you need to do your own research before using any images it produces based on your query. How different is searching it’s image index and producing potentially copyrighted images it found online, from copilot’s search of ‘how to code’ doing the same thing?

> Why can’t copilot also rely on that same ‘user must recognize that some of the results produced might be copyrighted’ logic?

Google shows you where you can find the copyright and license information. GitHub Copilot does not.

> Worth noting that Google image search has long presented its results under a caption pointing out that the results may be copyrighted and you need to do your own research before using any images it produces based on your query.

Because there's a common misconception that "Google images" are legal to use for any purpose. Afaik, Google has no obligation to correct people on this; nonetheless, they choose to do so.

Google search sends you to where it found the code.

That the copyright and license information are maybe available there is not something Google search actually knows to be the case.

That information might not actually be there, and Google does not care either way.

Similar to photos, There’s maybe a common misconception that code produced by copilot is free to use for any purpose. Is copilot under an obligation to point out that that might not be the case?

> That the copyright and license information are maybe available there is not something Google search actually knows to be the case.

Google sends you to where it found the code, on a best-effort basis. Google can't actually, in the general case, do any better than this. Google is a search engine; it's widely understood that, as a search engine, it looks for things and then tells you where it found them (not where they originated from; that's not what the verb "search" normally means, in English).

> Is copilot under an obligation to point out that that might not be the case?

GitHub Copilot is not branded as a search engine. It doesn't give people the option to identify, or verify, the terms under which the output it produces may be used. It doesn't provide any way of discovering the attribution, even in principle.

Normally, when something is given to you without any copyright markings, without any attribution of any kind, or licensing information, or a source:

• the computer probably came up with it on its own, just from the input you gave it; and

• you're free to use it.

This is not the case. GitHub Copilot goes against the expectations of the user; its UX design is actively misleading. This misconception only exists because Microsoft created it.

> You'll never know if somebody has lifted a bit of code in violation of a license anymore than you would be able to tell if somebody copy-pasted from stack overflow without attribution in any commercial application

Wasn't this the subject of millions of dollars of litigation between Oracle and Google?

For others like me, missing context:

> Google’s copying of so-called application programming interfaces from Oracle’s Java SE was an example of fair use, the court held (…).

Source: https://edition.cnn.com/2021/04/05/tech/google-oracle-suprem...

> Wasn't this the subject of millions of dollars of litigation between Oracle and Google?

That was a very different case. That case was about whether an API/interface was considered copyrightable. The outcome of the case was that the implementation was obviously copyrightable however the interface was not. The background being google creating their own custom implementation of the JVM.

The outcome (SCOTUS decision) did not decide whether the interface was copyrightable or not. It decided that if the interface were copyrightable, it'd be fair use anyway, so there's no need to bother resolving that question.

I was thinking this, but it is almost thr inverse. Api's without their implementation are not proprietary, but is an implementation without it's API (and surrounding structure) proprietary?

It's obvious that you've never endured any sort of deep code analysis for license "problems" because then you would know that it is in fact comically easy to tell if someone did this. It happens all the time.

While some detections are false positives, this is in fact why such deep analysis is performed: because getting caught selling a commerical application you don't have the rights to sell can destroy value very rapidly. It halts mergers and acquisitions. It gets products pulled off shelves. It gets people fired.

This doesn’t fit the definition of fair use, if that’s what you are implying. Fair use is a balance between four factors, transformative nature is just one factor.

To satisfy section 107 of the U.S. Copyright Act, you need to look at four factors. The first consideration is the purpose and nature of the work, which is where a determination is made as to whether a work is sufficiently transformative to be considered fair use of copyrighted material. There are other factors at play on this test, however. A court would look at what sort of commercial interest and benefit is being made from the use of the material, and even then they may decide that a non-profit motive is not sufficient to satisfy this test.

The second factor a court will look at is the nature of the copyrighted work. In the case of open source material, it is already published but most licenses quite rightly ask for attribution. A court may decide that not satisfying the attribution of the work may cause it fail this criteria, and prevent it from being fairly used. However on this point, I feel CoPilot may be on more steady ground as the work is already published.

Fair Use doctrine also factors in the amount and substantiality of the work being reproduced. It is the substantiality issue that will likely trip up copilot, as the code snippets in use may be complex enough in nature that it is considered a substantial part of the code. A court may well consider an algorithm that took time to create and test, although only a small part of the codebase, is sufficiently complex in that it took substantial effort to create and thus look dimly on its use in a larger derived work.

The fourth factor in play here is the one being argued by the litigant, and I feel will be extremely hard for GitHub to justify; that factor is the effect use of the code has on the potential market or value of the code. Open Source projects rely on volunteers and contributions. Substantial effort is made to actively maintain projects. It is not within the spirit of most projects to have people use unattributed code snippets, because it reduces the value of contributions and improvements to the code project. Improvements and contributions are reduced, and it is also hard to distribute improvements and fixes to any project that uses the code. In this regard, CoPilot seems to be completely violating Fair Use doctrine.

So as you can see, Fair Use is way, way more complex than most people realise. Many people think that a sufficiently transformative use of copyrighted material is sufficiently for satisfying a Fair Use defence, but in reality it is not.

Exactly. It's like people calling for Dall-E to be banned. Well Stable Diffusion exists, as do a dozen other similar projects and an entire ecosystem of training data and models. The march of technology rarely stops because of people crying "copyright!"

> "The march of technology"

Who are the stakeholders putting the machine to the run? Who owns the machine?

This is much more subjective than clinical rhetoric praising tech may allure to.

BigCode project is coming and it gonna be opensourced the same way stable diffusion is.

We can't expect a massive improvement on computer performance anymore. So it will never come the day household computer can finish training the model equivalent of Github Copilot in practical time.

They can just add more cores and more power. With modern languages like Rust making multi threading more accessible, I expect we will double down on this. You could also crowd source this, some distributed application where everyone puts their home machines towards training.

> You could also crowd source this, some distributed application where everyone puts their home machines towards training.

That's how Leela Chess 0 (LC0) replicated the Alpha-zero performance. In fact, this is actually not that difficult. Assuming you have the means to orchestrate it; all it takes is loading the weights and a batch, computing backprop; and submitting it to the central system which aggregates the gradient updates and then updates the whole network and push new weights (kinda how bitcoin creates a new block).

This is no different to gradient accumulation; just "distributed". In-fact, the system could offload a large number of batches because the returned update is O(1) space to return where n is for batch size; it's just that the O(1) is the size of the network.

I’d argue, running it on your local machine as your own copilot seems like a better proposition than what is offered now.

There is still the issue of setting boundaries on what to learn from, but it is much more aligned with you taking the responsibility for what your copilot does or doesn’t, and it might give a better window of opportunity to customize it to your needs.

We’d lose economies of scale, but also lose infringement at scale, and gain individual control. Net positives overall ?

Any ML model that decides to use my code as part of its code suggestion functionality has a fool for an algorithm.

5. Microsoft, a $1.5e15 company, will not profit from the technology (as easily).

If that is so, people will stop writing open source. Then this model will use its own output until it produces bland garbage like a camera filming its own output stream.

“People will stop writing open source” is far from being self-evident.

I write open source code to benefit others and to put more learning material in to the world. If I was so paranoid about other people using it I would have made it proprietary.

Linux has a lot of corporations contributing code to it. Tensorflow is backed by Google. Plenty of other projects are released as open source by companies, not people. I don't see companies crying about Copilot, so their employees will keep writing open source when instructed as such by their managers.

Reply on news.ycombinator.com