Legality of GitHub Copilot
theverge.com
theverge.com
I love the sheer fucking hubris of this statement. It's fair use because "the machine learning community" wants it to be. What a terrible justification. I hope they get their ass kicked in court over this.
Anyways, the problem here isn't that they're training on open-source code, it's that their LLM is outputting code that violates open-source licenses by not including the text of the license and the author's original copyright statement. That's not fair use and it never has been.
Keep in mind that this is a subsidiary of Microsoft doing this. I'm sure M$ won't mind my new "Machine Learning platform" that outputs a copy of Windows with a piracy crack. After all, it's widely accepted as fair use by "the machine learning community".
I as a human can go read the code, learn, get inspiration, write it in my own style, and be fine. What's the big difference you see between my learning and the automated process for the AI? Is it just that cross between wet and dry storage and "understanding" that causes the problem?
GitHub Copilot is a machine and the sum total of everything it knows is code from GitHub. Literally its entire universe! There is absolutely no "experience" it has ever had that is not code from GitHub. There is no other fount of knowledge or inspiration from which it can draw. It is incapable of doing anything other than rearranging the Lego blocks in its brain — in fact, it doesn't even have a notion of what it's doing!
Neural networks are a very complex math functions that take in some numbers, multiply them and spit out some other numbers. That's it.
My personal take - it's more like the infinite monkey theorem.
---
a monkey hitting keys at random on a typewriter keyboard for an infinite amount of time will almost surely type any given text, such as the complete works of William Shakespeare. In fact, the monkey would almost surely type every possible finite text an infinite number of times.
---
except for each letter, we're jostling the monkey's hand a little based on the letters he's already hit, and the training data.
Ex: There is a total problem space of "Valid Python" or "Valid JSON" and these machines are designed to explore parts of that problem space directed by the user and their training data.
And just like it only takes 52 cards to make any given shuffle essentially an entirely new & never before seen section of the problem space "decks of cards" (see: https://www.mcgill.ca/oss/article/did-you-know-infographics/......)
I don't really know of a clear way to distinguish what it's doing from "human inspiration & creativity" because those activities also tend to recombine previous experiences and data to create new but related works.
Basically - we also tend to explore related problem spaces based on what is currently happening around us and our life experiences (This is why the arts have period specific themes and styles).
This thing just does it a whole lot faster than we do. Whether that's good or bad is a different question.
This is core of your argument — "why isn't it the same?". I hear it so often and it's just utterly unconvincing to me. Neural networks are "function approximators" — that is, given a function f, you train a network on a subset of its domain and corresponding known values in its range, and you hope that when you feed in other x values you get accurate predictions for the y values.
I want to see an affirmative argument that this is how our brains work. Where is the evidence that we can be modeled as functions in this way? Why is your default hypothesis that neural networks are analogous to human brains?
I don't know why you think this matters.
I don't care at all if this is how our brains work - I care that the result is obviously analogous to what we call creativity - there are only so many ways you can arrange a 1kx1k pixel grid. Human digital artists explore that space, Machines can explore that space.
We happen to like some parts of that space more than others, and so we guide the machines to explore those parts more heavily.
So...
> Why is your default hypothesis that neural networks are analogous to human brains?
Because I can see that they are obviously capable of a similar skill.
Does it matter that your earth-mover is using hydraulics and not human muscles? Fuck no. It's digging.
Does it matter that this AI is using pre-trained weights and not human brain flesh? Fuck no. It's digging.
----
Now - entirely outside of that discussion: I also happen to think modern copyright is one of the absolute worst abuses in human history (and it's getting significantly worse right now, with forced rentals, digital oversight, locked down devices, DMCA, etc)
Our cultural heritage is OURS: collectively. I am not thrilled, but accept, the original intent of 7 years copyright with 7 years renewal.
I am absolutely, morally, ethically, opposed to modern copyright of "Life of author + 70 years."
It means that as adults, we cannot tell the stories we heard as children in our own voices. We cannot bridge our culture between generations. We have sold our soul to fucking mickey fucking mouse.
Fuck that bullshit. I'd rather steal it all than play that game.
How is this analogous to creativity? How are they capable of skills similar to humans? You keep saying these things are “obvious”, but you haven’t managed to explain them once.
Because that seems to be the only way to justify ip theft used for turning ml models into viable products. If this assumption is invalid then so is everything else - thus they push for it in a desperate attempt to productise other people’s work for free.
If ai would “think” and “learn” then surely they wouldn't need ever increasing amounts of data. Yet they do, because their premise is incorrect.
> If ai would “think” and “learn” then surely they wouldn't need ever increasing amounts of data. Yet they do, because their premise is incorrect.
If only people would "think" and "learn" and then surely they wouldn't need [education and basic input from the world]. Yet they do, because clearly people aren't thinking, learning creatures.... (fucking /s)
Like - I'm struggling to even find a charitable way to read your comment...
>What's the big difference you see between my learning and the automated process for the AI?
The root of the problem is that machine learning community wants AI to be treated as a human when that's convenient for them and as a machine when it's not.
I guess my question is, where does learning end and copying begin? And, does this apply the same for humans and AI?
>And, does this apply the same for humans and AI?
It shouldn't. For one thing, the process referred to as "training" would be more accurately described as "tuning parameters" and has no proven similarities to human cognition.
For another thing, we operate in what could be termed a "human-supremacist" legal system in which computer programs have no rights, so there absolutely could be laws and court rulings that are discriminatory against computer programs.
Those are mistakes and they will become more rare as the AI improves. If the system was designed to store chunks of code and output a copy of existing ones, the vast majority of output we see would be exact copies.
They are not just mistakes, they are mistakes at scale. Mistakes en masse. Thus, they are mistakes that should carry a heavy weight and a heavy culpability for the wreckless carelessness that brought them about.
If I made a mistake and my software enabled mass media piracy, lawyers can, would and should come after me regardless of me or my software's pledge to learn and make those mistakes "more rare".
Especially if I keep committing those mistakes and cover up for them by adding individual specific blacklists rather than making expensive systemic fixes.
This is what I was saying earlier about AI companies talking out of both sides of their mouth; when it's convenient for them they want AI to be treated as a human but whenever something goes wrong they talk about how it's not their fault, the system will get better at this, some problems are inevitable, etc. You don't get to have it both ways.
AI is not a human. It doesnt learn or get inspiration. It’s software that ingests data. There’s nothing human about it.
Even if it were, humans cant just steal tiny bits of other people’s work, reassemble it and call it something new. It would be like stealing a thousand paintings, cropping tiny bits out of each and making a “new” painting. It would still be made of stolen bits even if there are millions of such bits that mixed anew resemble none of the originals.
I mean look at Madeon’s Pop Culture song - wouldn’t “stealing tiny bits of other people’s work; reassembling it and calling it something new” describe it pretty completely? Except that track is a masterwork in the mashup genre. The bits aren’t ‘stolen’ they’re just ‘used’ and the entire concept of the song revolves around them being immediately recognizable.
What is your litmus test for "understanding" and "comprehension"?
I don't disagree that it's a derivative work. But, I think the majority of a humans day to day work is derivative.
I can look at a bunch of faces, consider my interactions with other faces IRL that aren't copyrighted or restricted and my output can be original because while I am informed of all the faces I have seen, the mona lisa I draw is an expression of thoughts and views not a mere modification based on my understanding of the concept of faces.
In other words, LLMs use a prompt to do stuff, they don't just say and do things because they want to express themsleves, therefore their output is not their expression or the expression of their owners but a processing of their input based on the prompt.
To simplify this even more, if a script parses the Linux Kernel's code and outputs some video game because the programmer told it to, and it's only input is the Linux kernel then is that video game subject to GPL?
If a human reads that same code and writes a video game, even if that human only knows the kernel's code we can say that since humans think and apply their own ideas, the kenrel code was merely informative.
Now, if GPT4 is able to use it's training data (internet) to output arbitrary content without being prompted and the results of a prompt are based on it's normal thought process that wis independent of a human prompt then you can argue it is like an assistant helping the prompter human, but right now it is a tool that "understands" (i disagree but that's a different debate) publicly available internet content and processes prompts based on that "understanding" without applying it's own ideas, feelings or thoughts. It can be argued that its output is a mere processing of largley copyrighted content, similar to re-encoding a copyrighted movie.
I would think what makes or breaks this in a US court is the question of "can a reasonable person who knows how chargpt works look at a prompt, output and conclude part of the output at least was a mere processing of, as opposed to original content that was inspired by the copy righted content?"
> "can a reasonable person who knows how chargpt works look at a prompt, output and conclude part of the output at least was a mere processing of, as opposed to original content that was inspired by the copy righted content?"
It would be funny if the "workaround" was to disallow very low "temperature" output, for some sort of "no, it's random, not just statistical!".
If the output of an AI is purely functional, there is no copyright concern as the AI did not use any copyrightable content in determining the output.
>as the AI did not use any copyrightable content in determining the output.
uuhhhhhh...didn't it? I thought we were talking about AI systems that are outputting code from their training sets verbatim. I don't see how you could possibly argue that the code it copies from its training set wasn't used to copy that same exact code from its training set.
I can certainly take GPLv3 code and plop it into my app if the code is not subject to copyright. Without the copyright, there is no license.
But one thing that seems almost universally true is that there is an incredible amount of uncompensated and undercompensated labor behind every success in this domain.
Which I see as simply a refinement of the overall model of capitalist advancement of technology. But it's very stark in this case, to the point where normal people are able to clearly intuit a wrong, even if they can't articulate it to the satisfaction of the advocates & beneficiaries.