Remix.run Logo
▲ jonhohle 11 hours ago

I reached out to both this legal team and SFC about cases where all of the LLMs will license strip open source code. In cases where BSD licensed code was copied into thousands of projects and millions of forks, it appears the training is so embedded that it will produce functions down to the formatting decisions. In the case of `strlcpy`, no human written version looks like the OpenBSD version, but most commercial LLMs will license strip the OpenBSD version when asked to implement. Depending on the context (not trying to trick them, just other programming questions), I’ve had them all produce it without license (100% score in JPlag).

▲lilbigdoot 11 hours ago | parent [-]

I keep wondering how this can of worms is going to play out. Right now it doesn't get them into trouble because they're taking from OSS communities, but what happens once one company finds out code they released with a restrictive license got copied verbatim into a closed product? Will they retaliate? Will they decide together to not open pandora's box? I wonder really

▲kodablah 8 hours ago | parent | next [-]

Anthropic and OpenAI offer indemnity for this situation for their enterprise customers if such a suit were to happen. They will defend you and cover costs.

▲spwa4 11 hours ago | parent | prev [-]

It doesn't matter. The courts are entirely unfair. Take the copyright office's arguments: it states the court should explicitly address "Market Harm". Essentially, ANY fair use, such as AI training should be illegal, if it harms the market for the original authors EVEN WHERE THE LICENSE WOULD NORMALLY ALLOW THAT USE. That was reasoning applied to protect billionaire profits from the horror of library lending. The copyright office's position documents this extensively.

https://www.copyright.gov/ai/Copyright-and-Artificial-Intell...

See IV. Fair use, D. Factor Four

In other words, no matter the license, the fact that OpenAIs fair use of the works for training MIGHT take opportunities from the plaintiffs should, BY ITSELF, be an argument to make that AI training illegal. Even if the software programmers agreed to allow AI training, then still sued OpenAI, that is a valid argument accepted from billionaire companies.

In case you're wondering how the court argued that in this decision? It didn't. Not covered at all. Obviously brought up by the plaintiffs, not adressed. The weirdest pro-billionaire arguments are just accepted by the court (the big one is accepting AI models as making new works despite the copyright act clearly stating, 50+ times, only "natural persons" can make new works), and the most reasonable, long established principles, such as market harm (if the use of a copyrighted work harms its authors in the smallest way, even lost sales, even lost job opportunities, lost licensing opportunities, ...) is enough to forbid it's use EVEN WHERE THE LICENSE STATES DIFFERENTLY.

This was required because they wanted to forbid digital lending out of Disney's movies, where the lending software would guarantee that only one copy is actually in circulation (ie. use of DRM to let a library just show a movie without the patron coming to the library to get it). In other words: when billionaires want to make money of works, and the first sale doctrine as well as long standing fair use principles for libraries were in the way, the courts declared that market harm superseded fair use. And, of course, suddenly, when it protects programmers from billionaires, it doesn't apply anymore.

The courts deciding this way has done incredible damage to the availability of older works, to the point of destroying them entirely, since libraries cannot legally keep collections of older works and make those available over the internet, and it is impossible to keep the physical media in working condition and available over a large area, like the united states, for example.

Repeating theme here: absurd arguments to protect billionaire profits, silently rejecting and ignoring reasonable arguments to protect normal people.

▲jonhohle 7 hours ago | parent [-]

I looked into this a few years ago for lending older games or unique prototypes or press copies to researchers without having to send them a copy. VGHF argued for this use case with the copyright office and was denied an exemption to the DMCA, as the use case was not considered fair use (which affects DRM free media as well which is not affected by DMCA).

▲spwa4 7 hours ago | parent [-]

Given what happens if you ask Claude to reproduce some well-known game, especially FPS games ... you can be pretty damn sure it's seen the game, maybe they even trained on all the leaked source code they could find. But I find it very hard to believe that style, that code, the best game styles came from open source code.

I mean if it trained on open source, you'd expect it to always come up with Quake essentially. That's not true at all. It's always very close to Call Of Duty in how they play.

Very hard to believe in coincidences.

90% sure Anthropic did just that.