Hacker News Reader

Comment by jedberg | Hacker News Reader

No surprise there, given that Japan has declared that they will not enforce any copyright claims against AI companies, and in fact explicitly allow training on copyright material, even stolen copyright material.

https://www.privacyworld.blog/2024/03/japans-new-draft-guide...

View on HN

It would be remarkable if the software world were to suddenly depart the US, because a temporarily-hostile regulatory climate made certain startups migrate elsewhere, and they became wildly successful there. I don't think it's crazy to worry about that happening here and now, is it?

(Sure: the US government seems to act very friendly to one startup, OpenAI—but that's not the same as being friendly to OpenAI's competitors—it's sort of the opposite. OpenAI creating regulatory moats in the US is also OpenAI making the US a comparatively unattractive jurisdiction for everyone else..)

Tech specifically is dependent on venture capital because it requires upfront costs and produces something that is not a valuable asset in liquidation. This basically guarantees that tech will go somewhere with large liquid capital markets... of which the US is the largest by huge margin.

That said, there is ample indication that companies will continue to outsource to foreign talent (eg. Google laying off California workers and hiring Indians). Between outsourcing and remote work, it's increasingly possible to start a business in one location, but employee people globally.

I think the real outcome, which we're already seeing, is that companies pay for their training content. It fully integrates into the existing venture system ("simply" raise more money), integrates into existing laws, and feeds into the existing tech economy.

In order for the startup ecosystem to leave the US, there would need to be either free-flowing checks rivaling the US or a dramatic regulatory change. Japan just doesn't have either, even with this new policy.

Not to mention local talent and experience pools, which are what almost everyone overlooks when they talk about a "new silicon valley" popping up someplace else, be that Berlin or London or whatever. Those places simply don't have the people or the culture to build a thriving tech scene at the same scale and velocity that SV does.

The question of whether training is a copyright violation is still an open one.

I think Japan is in the right here. IMO, training an AI is analogous to reading a book. As long as the AI isn’t regurgitating the contents 1:1, it’s not infringement.

Yes, I've been saying this. It's like if I invented a robot that can run, and then suddenly people started saying it shouldn't be allowed to run down the sidewalk. What's the difference between a human running down the sidewalk and a robot running down the sidewalk? I guess someone could argue that the robot runs a million times faster than any human can, and that sidewalks weren't built with this in mind, but I personally don't see a difference since my robot is just "running" and humans can run too.

For running, you’d probably be right. No copyright on running.

But for copyrighted content, it’s a thorny issue. You already cannot do whatever you want when you buy a book (or a cd, or a dvd); it’s for personal use unless you pay more. Surely you can quote some paragraphs or take a couple of screenshots, and you can learn the overall content, but you cannot reuse it in its entirety.

Also, the usage of those contents is kind of “rate limited” by the fact that an human must spit out those answers.

Now with AI you get a kind of an opaque system which “ingests” data in some way that can be considered “compression” (rather than “summarization” or “understanding”) and can spit this out at very large rates. Is this a fair thing? If you got a small model, probably your model has learned the concept, so I could agree that copyright is not a concern; but with LLMs? The NYY lawsuit is very relevant.

Also, I suspect that humans will adapt and create content which is less useful or harder to digest for AIs - maybe by making it worse for humans as well. Or they will use AIs to create bad content on their topics of expertise to make AI training data worse. There’s a serious risk of horrible content flooding the internet (and the publishing world).

LLMs are useful but risky. We should take care and consider 2nd order consequences.

I personally wouldn't want a robot running a million times faster than a human down the side walk. The argument that it's running just like a human runs isn't convincing to me.

Honestly? Pretty smart move by them. Japan's major export is its culture. Nowadays almost everyone enjoy manga and anime and by extension, they adore Japan. That kind of soft power can be very useful. By allowing unlimited access to their cultural outputs, including copyrighted materials, they can influence the AI and further boost their reach.

You can sort of see this in Stable Diffusion and music models like Suno. Lots of creations are basically Japanese art.

How does that work then if a model trained on that data is used outside of Japan and some rights holder makes a claim then (in another country)?

  In teasing out this apparent dichotomy, the committee first focuses on distinguishing between digesting copyrighted content for “information analysis” (which is allowed) versus the use for a “purpose of enjoyment of the thoughts and or sentiments expressed in the copyrighted work” (which is not allowed). The committee’s report states that “enjoyment” consists of satisfying the intellectual and mental needs of the individual through the viewing/experiencing the works. By contrast, information analysis might include a scholarly assessment of movie themes across various genres, while “enjoyment” would likely include the ability to play and view all or a significant part of the individual movie.

Interesting, I wonder how they will enforce it. I could envision a world where AI datasets become split between commercial / non-commercial, but this sounds like the dividing line is whether or not it the end work is "enjoyed".

What if GPT "enjoys" reading something lol

And just like that Japanese replaced english as the universal language in the 21st century, it just took a generation to learn the best language for interacting with machines. <some historian, 22nd century, translated from Japanese>

When we look at this new legistration, together with other copy right policy Japan enforce [0], and how most of the Japan website think disabling copy and paste is a good idea to maintain copyright. One will be very puzzle what they have in their mind.

[0]: https://www.japantimes.co.jp/news/2024/04/18/japan/crime-leg...

And yet Nintendo sues anything that looks even a little like Mario. This should be interesting to watch play out.

According to the link I believe they'll still be able to sue and win.

You can train a model, but as for using it for commercial use that's a different story.

Reply on news.ycombinator.com