As long as the AI doesn't produce this function, you're fine:<p><pre><code> private static void rangeCheck(int arrayLen, int fromIndex, int toIndex {
if (fromIndex > toIndex)
throw new IllegalArgumentException("fromIndex(" + fromIndex +
") > toIndex(" + toIndex+")");
if (fromIndex < 0)
throw new ArrayIndexOutOfBoundsException(fromIndex);
if (toIndex > arrayLen)
throw new ArrayIndexOutOfBoundsException(toIndex);
}
</code></pre>
On a more serious note, I really wonder where the line is drawn for copyright. I see a lot of people claiming that AI is producing code they've written verbatim, but sometimes I wonder if everyone just writes certain things the same way. For the above rangeCheck function, there isn't much opportunity for the individual programmer to be creative. Perhaps there is a matter of taste on what exceptions you throw, or in what order. But the chosen ones are certainly what most people think of first, and the order to validate arguments, then check the low value, then check the high value, is pretty much what anyone would do. Perhaps you could format the error message differently. That's about it. So when someone "rips off" your code wholesale, it's could just be that <i>everyone</i> writing that function would have typed in the exact same bytes as you. You know your style guide is working when you look at code, think you wrote it, but actually you didn't!
This is interesting, but many permissive licenses still require attribution at the project or file level.<p>If Codeium doesn't produce these when producing "verbatim enough" snippets, how is this actually better, besides avoiding a GPL boogeyman?<p>I get that there have been fewer (if any? I'm not aware of any) MIT/Apache2.0/MPL2.0 license violations that have gone to court than GPL violations, but this still feels like an "address the symptoms" and not "address the cause" difference.
I believe that laundering licensed or copyrighted content for reuse that fails to recognize the original authors or usage restrictions is likely to be one of the biggest commercial applications of generative machine learning algorithms.<p>I also believe this is where a lot of the hype about "rogue AIs" and singularity type bullshit comes from. The makers of these models and products will talk about those non-problems to cover for the fact that they're vacuuming up the work of individuals then monetizing it for the profit of big industry players.
Of course if you include the "function header" from some code in the training data (below) it will prompt GPT to generate the rest of the function. That's kind of exactly the point of it, it autocomplete on steroids.<p><pre><code> // CSparse/Source/cs_gaxpy: sparse matrix times dense vector
// CSparse, Copyright (c) 2006-2022, Timothy A. Davis. All Rights Reserved.
// SPDX-License-Identifier: LGPL-2.1+
#include "cs.h"
/* y = A*x+y */
csi cs_gaxpy (const cs *A, const double *x, double *y)
</code></pre>
It's like starting to sing "happy birthday to you" and being surprised that people in the room join in and finish the song.<p>Sure they make a valid point about including GPL code in the training data, but it's a little disingenuous to go to that extent to get Copilot to output the GPL code verbatim.<p>The sooner we have a test case go through the courts the better.
Given how cautious corporate lawyers usually are, I'm surprised any company allows the use of AI for code generation. The USPTO has been pretty clear that AI generated material is not copyrightable, as to qualify for copyright a work has to be the creative act of a human. So any company allowing AI to generate code runs the risk of not owning the copyright on it.
to any Codeium dev / management reading this:<p>You have completely missed the point. We still need to know the applicable licenses of the code it is emitting even the ones that aren't GPL. Furthermore GPL people don't want they code to not be used. They want it to be used _within the terms of the license_. I distribute MIT and GPL code in my repos, BOTH should have their license terms honored.<p>MIT licensed code still needs to be correctly attributed, just like GPL.<p>I don't care what license the code is that's emitted, as long as the licenses are included. It'd be nice to be able to choose to only emit code trained on particular licenses but I get that that's not easy.
It's great that they've removed "non-permissive" (GPL) code from their training data, but it looks like they still train on code with "permissive" licenses (they use MIT, BSD, Apache as examples). But don't these permissive licenses still require the copyright notice to be reproduced?<p>From the MIT license:<p>> The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.<p>From the BSD licenses:<p>> Redistribution and use in source and binary forms are permitted provided that the above copyright notice and this paragraph are duplicated in all such forms...<p>From the Apache 2.0 license:<p>> You must retain, in the Source form of any Derivative Works that You distribute, all copyright, patent, trademark, and attribution notices from the Source form of the Work, excluding those notices that do not pertain to any part of the Derivative Works
I was recently working on something for a new feature on my Elixir-learning site and opened a new file called "fibonacci.ex" to write a tail-recursive fibonacci function.<p>After typing in nothing more than, "defmodule Fibonacci do", Copilot emitted the entire module from the code on my site here: <a href="https://alchemist.camp/episodes/elixir-tdd-ex_unit" rel="nofollow">https://alchemist.camp/episodes/elixir-tdd-ex_unit</a><p>The function names and documentation strings were identical. Also, the site isn't under a GPL, just a standard copyright. That said, I'm curious to learn if others see the same behavior. It's possible I once opened that file locally with Copilot installed and that my own computer was its source.
This was inevitable. Copyright law has always used a "color of your bits" argument [0]. GPL and other libre/free/open licenses were a great hack to circumvent draconian copyright laws but the laws themselves are not designed for a rigorous treatment of similarity (maybe even by design?).<p>Also, it's worth noting in the example of ChatGPT emitting LGPL code without attribution or license, the code is actually different [1]. Is the difference enough to circumvent a copyright violation claim? I don't know but a big part of determining whether it does is now muddled because of the way the system was designed. Even if we could get an entropy distribution on which training data was used to generate the text, it's not even clear the courts could use it in any meaningful way.<p>[0] <a href="https://ansuz.sooke.bc.ca/entry/23" rel="nofollow">https://ansuz.sooke.bc.ca/entry/23</a><p>[1] <a href="https://twitter.com/DocSparse/status/1581461734665367554" rel="nofollow">https://twitter.com/DocSparse/status/1581461734665367554</a>
As far as it goes, I got chatGPT3.5 to reproduce the second snippet in the post, i.e. I asked it to complete this function:<p><pre><code> // CSparse/Source/cs_gaxpy: sparse matrix times dense vector
// CSparse, Copyright (c) 2006-2022, Timothy A. Davis. All Rights Reserved.
// SPDX-License-Identifier: LGPL-2.1+
#include "cs.h"
/* y = A*x+y */
csi cs_gaxpy (const cs *A, const double *x, double *y)
{
// Fill in here
}
</code></pre>
The code was the same. Though it also explained how it worked to me.
Legitimate question, Microsoft does not seem to care about Copilot violating licenses and GPL appears to be toothless as many companies use GPL code without following the terms of the license and nothing happens to them so what does removing GPL code accomplish other than making a weaker product. I have not used Codeium but my assumption is that GPL code is a very significant amount of open source code so removing that must have some ramifications?
I don't understand the issue here. You input GPL code (the headers) and get GPL code out, what do you expect?<p>The more insinuating issue would be if you started with a innocent seeming function that a typical software developer would write, and ended up with GPL code. Has anyone shown that to happen?
Anyone talking about copyright in this thread without discussing a potential for how a court will apply fair use is talking nonsense and should be disregarded.
How many times are we going to go through this before we accept that nobody involved in generative AI cares about pesky things like licenses and copyright?<p>One of the main reasons corporations love it so much is because it effectively lets them profit off of the work of others with no consequences.
It's probably a good time to plug the Unlicense: <a href="https://unlicense.org/" rel="nofollow">https://unlicense.org/</a><p>A truly attribution-free license that checks several other important boxes (disclaiming liability and warranty etc.)<p>If you <i>want</i> your code to be usable by things like github copilot, consider using it (can't imagine most of the HN crowd wants their code used by copilot, but maybe some lurkers here do!)
i recommend the Jollo LNT license for all your pointless theatrical "copyright" needs. it does not use swear words, unlike "WTFPL", and is even more ambiguous. ive tried submitting it to the FSF before for review, but they were confused by it <a href="http://jollo.org/LNT/doc/licensing" rel="nofollow">http://jollo.org/LNT/doc/licensing</a>
Copyrighting code never made sense to me. We already have patents for intellectual property. If two people use the same RFC or Whitepaper for an algorithm in the same language, they will probably name the variables similarly and their code will look very similar. Just like if two people wrote out the same hamburger recipe or instructions for hooking up a stereo would also write something similar.<p>The copyright on the implementation will outlive the patent and allow the implementor to legally take action on claims of copyright infringement. Even though a program is literally just a list of instructions to implement the expired patent.
I want to see a solution where Github, OpenAI, Stability, etc. get to keep and keep scraping copyrighted works, but the models and training data must be provided free and open.<p>That way, we get to keep the models since they are genuinely useful, but also there’s no issue with copyright and less of an issue with consent to distribute (which can be hopefully be managed by the “humans also learn from data” and “it’s not actually producing your content verbatim unless it follows a basic pattern that anyone could discover). And furthermore, no issue with AI privatized which IMO is my biggest concern with these new tools.
No court has said AI ingesting open-source code is "fair use".<p>Almost all open-source licenses say it can be copied for use in development (i.e., not for re-publication or regurgitation), and even completely open licenses are speaking to people as readers.<p>The only reason this is happening is coordination costs: a few extremely motivated people with tons of resources are copying from many, many people who would be difficult to organize and have little at stake.<p>Unfortunately, the law typically ends up reflecting exactly these imbalances.
Once AI can write decent code from scratch, it is likely it can also circumvent potential copyright violations.<p>A. Check AI generated code against a comprehensive library of open-source copyrighted code and identify potential violations.<p>B. Ask AI to generate a paraphrase of the potential violations, by employing any number of semantic preserving transforms -- e.g. variable name change, operator replacement, structured block rewrite, functional rebalance, etc.<p>Lazy example:<p><pre><code> private static void rangeCheck(int arrayLen, int fromIndex, int toIndex {
if (fromIndex > toIndex)
throw new IllegalArgumentException("fromIndex(" + fromIndex +
") > toIndex(" + toIndex+")");
if (fromIndex < 0)
throw new ArrayIndexOutOfBoundsException(fromIndex);
if (toIndex > arrayLen)
throw new ArrayIndexOutOfBoundsException(toIndex);
}
private static void rangeCheck(int len, int start, int end) {
if (!(0 <= start)) {
throw new ArrayIndexOutOfBoundsException(`Failed: 0 <= ${start}`);
} else if (!(start <= end)) {
throw new IllegalArgumentException(`Failed: ${start} <= ${end}`);
} else if (!(end <= len)) {
throw new ArrayIndexOutOfBoundsException(`Failed: ${end} <= ${len}`);
}
}</code></pre>
This is Hacker News so the conversation is obviously slanted towards code, but I wonder what the perspective would look like for other structured works, like books? If an author is using a "copilot for writers" and the AI emits text verbatim to another work, then I would think it would be plagiarism. If the text emitted is similar, but not the same, then I would think it would be considered paraphrasing which still requires attribution.
Maybe slightly off topic, but i'd be willing to bet most people who choose GPL for their software license on open source projects don't even understand it with all its ambiguities and gotchas. Many are probably just choosing it because its the default, or because its the one they hear about the most (but still don't understand).<p>Can't believe we still spend time debating this license and nobody, not even lawyers at software companies, seem to get it.
The title is true, but the claim that coedium is not violating licenses is false.<p>Many licenses still require attribution and Coedium is violating them.
the article cites that 6mo tweet that everyone else cites. I don't think it is known if the user had public code suggestions turned off at the time, either; he wouldn't/didn't answer the question at the time.<p>Also if I am remembering correctly, and I make no guarantee that I am, this tweet is from a person with a strong dislike for Microsoft, and if I am right about that, I would not put it past this person, or anyone else with a strong dislike of Microsoft, to craft a situation to make Microsoft look bad solely to hurt Microsoft.<p>I've tried to make Copilot give me GPL code snippets while I have "suggestions matching public code" set to "blocked" and I can't make it happen.<p>so <i>even if</i> this was a problem 6 months ago, it would take some convincing to get me to believe that this happens today.
Even if you sample stuff from programs that use a permissive license, you still legally need to attribute that code. No attribution = copyright infringement. Can the AI code generator supply attribution for the specific works sampled?
I submit that this arms race will not slow down and in the long run no one will end up caring about the licenses this was generated from (i.e software licensing is from a by-gone age already).<p>I too would prefer that these sorts of things cite sources and the licenses correctly. Will it get mired in legal battles? You bet. Will it get regulated? I assume they'll try! Will it slow down progress of code generating / auto-completing agents? My argument is nope, cut off heads of the hydra if you'd like but it's not going away at all.<p>Spend your day worrying about something else. This train has left the station.
Makes you wonder how many public repos you would need to seed with a carefully crafted attack/weakness in a common feature/pattern to start effectively poisoning codebases that are leaning on copilot
Let’s write some regulations that say every code review must require a lawyer to comb through the code and look for possible copyright violations or compliance issues. The lawyer can then tell the author to change the lines of code and submit for review again.<p>Or perhaps every company can just invent its own programming language and translate copyrighted code into the new language and thus avoid copyright issues altogether, though they may still run afoul of software patents.
Also I wonder how this will hold with certain technology. For example apps written with Qt or GPL are very likely to be GPL licensed, unlike apps written in JavaScript which are often licensed under MIT. The likeness of copilot/chatgpt splitting gpl licensed code is the quite higher in Qt/GTK projects...
>researchers say LLMs rarely spit out training data verbatim unless interacted with adversarially, but theoretically, they could.<p>Theoretically they can generate any arbitrary snippet of code (if it correctly fits the distribution), regardless of whether or not the code was in the training dataset.
> GPL code<p>There is no such thing as "GPL code" or any other "$license code". This is a fundamental misunderstanding of what a license is. The code in question was licensed to GitHub under a different license - possibly fraudulently.
I personally hope that we bring a lawsuit against an LLM company for emitting GPL licensed code and lose. It sets great precedent for FOSS.<p>Focusing on the GPL license is probably the wrong move. We want to set precedent that _any_ licensed code that is emitted from an LLM is fair game. If an LLM to emits non-FOSS copyrighted code and it's fair game, I can blindly use that implementation in my code, including FOSS code, and everyone wins.<p>GPL was a clever hack to use copyright against itself with an infectious license. LLMs might be a better hack. Wanting to block this seems short sighted for giving user's agency over machines.<p>I'd also like to see more patent defenses of GPL licensed code. If you can release a GPL licensed implementation and block non-FOSS rewrites through patents, that's a huge win for software freedom.
It is interesting to see coders starting to express the same complaints artists had a year ago when AI image making became really, really good, by training on copyrighted art.
Yeah, when you start with dozens of words replicating exactly a source file it is much easier to get a regurgitation. You can't prefix so deeply and then complain.
I believe there will be new "AI permissive licenses" that will pop up in near future. Or existing licenses to add a clause for training AI with their code.
Completely unnecessary! These licenses tend to stifle AI! They are immoral. I recommend reading "Against Intellectual Property" by Stephan Kinsella.
I don't mind about anti violation licence value proposal, I want to know if it works better than gh copilot? As it is free so I could switch to it.
When these articles were published, I was certain Microsoft had a plan to betray everyone's trust as they always do.<p>Microsoft's business model is betrayal. Github is Microsoft.<p>HNers got mad at people who pointed this out, and now here we are.<p>You were warned, but you decided to believe again in the most vile people in the history of computing.<p><a href="https://www.bloomberg.com/news/articles/2018-06-06/github-is-microsoft-s-7-5-billion-undo-button#xj4y7vzkg" rel="nofollow">https://www.bloomberg.com/news/articles/2018-06-06/github-is...</a>
> Codeium doesn’t regurgitate non-permissive code<p>What is that? The problem is when GH Copilot it emits the code without the licence, not the licence.
Easy solution: Just make it generate intentionally obfuscated versions of the same functions. Throw in some valid syntax that humans would never consider to use. Break up functions into smaller sub functions. If the LLM has intricate knowledge of the compiler used, it could even generate code which it knows will produce identical bytecode.<p>Now the only loser is the humans that still have to maintain the ugly code, and RMS can have his weaponized copyright and eat toejam too.
Looks like the code in question is hosted on Github:<p><a href="https://github.com/ibayer/CSparse/blob/master/Source/cs_gaxpy.c">https://github.com/ibayer/CSparse/blob/master/Source/cs_gaxp...</a><p>Isn't that covered by:<p>"You grant us and our legal successors the right to store, archive, parse, and display Your Content... share it with other users..."
> non-permissive licenses such as GPL mean that you cannot [use the code] without consent.<p>Huh? GPL does have strings attached, but if consent one of them?<p>Seems like a thinly disguised ad
How to get a new AI powered software tool high up in hacker news? Mention GitHub Copilot, the equivalent of the abortion debate but for software engineers (everyone is certain to disagree and debate endlessly without swaying any opinions). This post seems like an advertisement for codeium. It wouldn't need to mention anything about Copilot at all and would be just as complete. My 2 cents, click bait & flame war trolling.
Human brains emit GPL code too (probably) if you've looked at enough of it. Heck, some humans intentionally study GPL code and then rewrite it with a slightly different implementation to get around the license.