Hacker News Reader

Comment by 6gvONxR4sf7o | Hacker News Reader

It's always weird to see the contrast between HN's reaction to copyright questions about text/image generation, and HN's reaction when it's code generation.

When a model is trained on 'all-rights-reserved' content like most image datasets, the community say it's fair game. But when it's 'just-a-few-rights-reserved' content like GPL code, apparently the community says that crosses a line?

Realistically, this tells me that we need ways for people to share things along the lines of all the open-source licenses we see.

You could imagine a GPL-like license being really good for the community/ecosystem: "If you train on this content, you have to release the model."

View on HN

A) This is just taking divided opinion and treating it like a person with a contradictory opinion (as others have noted).

B) Nothing about GPL makes it "less copyrighted". Acting like a commercial copyright is "stronger" because it doesn't immediately grant certain uses is false and needs to be challenged whenever the claim is made.

C) If anything, I suspect image generation is going to be practically more problematic for users - you'll be exhibiting a result that might be very similar to the copyrighted training, might contain a watermark, etc. If you paste a big piece of GPL'd copy into your commercial source code, it won't necessarily be obvious once the thing is compiled (though there are whistleblowers and etc, don't do it).

> Nothing about GPL makes it "less copyrighted". Acting like a commercial copyright is "stronger" because it doesn't immediately grant certain uses is false and needs to be challenged whenever the claim is made.

GPL says "you have a license to use it if you do XYZ." The alternative is "you have no license to use it." How is that not strictly "stronger?"

The GPL is as strong as a commercial license in the sense that the conditions it does specify are exactly as legally binding as those of a commercial license.

Right, but a license is more permissive than no license.

If this was a website for artists instead of programmers you'd see the exact opposite pattern. Unfortunately, people only seem to care when it threatens their own livelihood, not when it threatens that of the people around them.

The average HN denizen gets so pissed off when I ask them to save their comment rejoicing in the inevitability of the elimination of my entire field, and take it out when next year's descendant of CodePilot gives them the same horrible sinking feeling that these things give me.

If Copilot can replace my job, I think that job should be replaced by copilot. I don't think that saving jobs should be a reason to not hinder progress. I hope that what I contribute to my company is more than whatever future version of Copilot can create, but if not I will try and find another career.

Will I be sad if I lose my job to AI/ML? Yeah, probably, but at some fundamental level that's why I've always tried to keep myself up to date with stuff that's harder to automate.

I don't care either way. If an AI can do your programmer job, it means you aren't using your brain enough.

An AI can probably do half of my day job because it's stupidly repetitive. Leadership imposes old ways, and they dismiss anything "new" (i.e. newer than 2005). For example, writing all these high-level data pipelines and even web backends in C++ despite having no special performance need for it. Even though I'm not literally copy-pasting code, I'm copy-pasting something in my mind only a little higher-level than that, then relying on some procedural and muscle memory to pump it out. It's a skill that anyone can learn, just takes time. If I didn't have side projects, I'd forget what it's like to think about my code.

Some old-school programmers complain about kids with high-level languages doing their job more efficiently, so they work on lower-level stuff instead. It's been that way for decades. Now AI is knocking on that door. But before AI, C was the high-level thing, and we got compiler optimizations obviating much of the need for asm expertise, undoubtedly pissing off some who really invested in that skillset. If I'm working on something needing the performance guarantees of C or Asm, and the computer can assist me, I'm all for it. Please take this repetitive job so I can use my brain instead.

And the copyright thing is just an excuse. Programmers usually don't give a darn about copyright other than being legally obligated to comply with it. So much of programming is copy-paste. GPL had its day, and it makes less and less sense as services take over. GPL locks small-time devs out of including the code in a for-profit project but does nothing to stop big corps from using it in a SaaS. The biggest irony is how Microsoft not only uses GPL'd code for profit but also ships WSL, all legally.

People don't like being told no.

The vast majority of all-rights-reserved content is either not licensable, or not licensable at a price that anyone would be willing to pay or can afford. Ergo we[0] would much rather see more opportunities to use the work without needing permission, because we will never have permission.

When getting permission is reasonable then people are willing to defend the system. And code is much more likely to be licensable than art.

I still think the "Copilot is GPL evasion" argument is bad, though.

[0] As in the average HN user

That's a good way to look at it: The difference between "no" and "yes if XYZ."

Maybe it's that people respect "yes if XYZ" more than "no" because there's some path to yes that way. In that case, it really does speak to the need for some open-ish text and image licenses, like "you can use my image in a model if you share the model with me."

> When a model is trained on 'all-rights-reserved' content like most image datasets, the community say it's fair game. But when it's 'just-a-few-rights-reserved' content like GPL code, apparently the community says that crosses a line?

I don't think this is right. I think different people have different views, and you're just assuming that the same people have contradictory views.

I'm not assuming that, but I see how it could read that way, since I'm being fast and loose with the language. The community (anthropomorphizing the blob again) definitely empirically reacts very differently to the two topics.

To me the difference is this: https://news.ycombinator.com/item?id=27710287

It's possible for generation models to perfectly memorize and reproduce training data, at which point I view it as a sort of indexed slightly-lossy compression, but it's almost never happening with image generation because the models are too small to memorize billions of pictures, it can't produce copies.

Stable diffusion 1.4 has been shrunk to around 4.3GB, and has around 900 million parameters.

I don't know how big Copilot is, but a relatively recently released 20 billion parameter language model is over 40GB. ( https://huggingface.co/EleutherAI/gpt-neox-20b/tree/main ) GPT-3, according to OpenAI, is 175 billion parameters.

It's possible there are some images in there you can pull out exactly as is from the training data, if they were to appear enough times, like I suspect the Mona Lisa could be almost identically reconstructed, but it would take a lot of random generation. I'm trying it now and most of the images are cropped, colors blown out, wrong number of hands or fingers, eyes are wrong, etc.

that's bc hn has a lot of gpl zealots who want special rules because they believe a viral license is a "fundamental good" and closing stuff off via copyright is the opposite. i don't agree and think viral licenses suck, but it's not a unpopular opinion on here.

Is there a difference here?

With code, you literally copy it and put it in a device and fail to provide the source, breaking the GPL.

With a copyrighted set of images, you scan those and break them down into some set of data and then never need to actually copy the images themselves -- my understanding anyway.

Does that set of data contain copied works? Or is it just a set of notes about the works that have been viewed by the AI?

The difference seems like semantics. You're taking compyrighted data, encoding it, then deriving a work from the output of that encoded data.

If I compress copyrighted works and redistribute them as my own I'm technically breaking them down into some set of data and never actually distributing copyrighted work, right?

> You could imagine a GPL-like license being really good for the community/ecosystem: "If you train on this content, you have to release the model."

The issue with that is that it is perfectly legal to ignore the license, if you use the copyrighted work in a transformative way.

It doesn't matter what the license says, if it is legal to ignore the license.

That seems to be the status quo, but if huge companies benefit from the electorate's work and use that to put them out of jobs, I wouldn't be surprised if the law changes.

There's no community consensus on either of them, so not sure how you are trying to draw a contrast. There are people who thing both are wrong and those that think both are right and everything in between.

Reply on news.ycombinator.com