Conversation

Notices

1. Embed this notice
   TheStrangelet :bc: (thestrangelet@beige.party)'s status on Tuesday, 25-Aug-2026 01:26:53 JST TheStrangelet :bc:

   • millennial fulcrum

   @falcennial Hostile... But, a large language model doesn't have to be trained on stolen works of art. A LLM doesn't have to used as an excuse to abuse the labor force. A LLM doesn't have to be used as a data surveillance vector. LLM companies don't have to constantly train more complex models using up huge amounts of energy and water. LLMs don't have to be used to target human beings in war. LLMs don't have to be used to poison the commons data-space. All of these things are choices wealthy, powerful people have made. These are choices.

   • Embed this notice
     Rich Felker (dalias@hachyderm.io)'s status on Tuesday, 25-Aug-2026 01:26:51 JST Rich Felker

     in reply to

     • millennial fulcrum

     @thestrangelet @falcennial It does have to be trained on stolen works. There is no way to organize the volume of information needed for "largeness" in this sense that can also ensure provenance. Even if you want to do it by distributing the work among millions of humans, there is no way to ensure that that many humans are playing by the rules. There are certain scales beyond which these things are unknowable due to fundamental issues of trust and epistemology.

   • Embed this notice
     Rich Felker (dalias@hachyderm.io)'s status on Tuesday, 25-Aug-2026 03:57:37 JST Rich Felker

     in reply to

     • millennial fulcrum

     @thestrangelet @falcennial The publishers do not have the rights to sublicense works they published for unlimited use in derivative works. For many older works the publishers have no rights at all; they reverted to authors or authors' estates.

     In general authors are not fond of their words being taken for the purpose of mimicking their writing style, and will not license it.

     A corpus consisting of books whose authors granted permission for them to be used is orders of magnitude smaller than what would constitute a "large" language model.

   • Embed this notice
     TheStrangelet :bc: (thestrangelet@beige.party)'s status on Tuesday, 25-Aug-2026 03:57:39 JST TheStrangelet :bc:

     in reply to

     • Rich Felker
     • millennial fulcrum

     @dalias @falcennial Are you claiming companies that are literally raising trillions of dollars can't afford to pay publishers for rights to use material for training?

   • Embed this notice
     Rich Felker (dalias@hachyderm.io)'s status on Tuesday, 25-Aug-2026 04:18:00 JST Rich Felker

     in reply to

     • millennial fulcrum

     @thestrangelet @falcennial You are moving the goalposts.

     First you said "doesn't have to be trained on stolen works of art".

     Now you're declaring that stealing any "publicly shared text on the internet" is "free game whether we like it or not".

     No, that is a corpus of stolen art and proof of my thesis that there is no such thing as a large language model not built on stolen works.

   • Embed this notice
     TheStrangelet :bc: (thestrangelet@beige.party)'s status on Tuesday, 25-Aug-2026 04:18:02 JST TheStrangelet :bc:

     in reply to

     • Rich Felker
     • millennial fulcrum

     @dalias @falcennial Older works that have expired copyrights and publicly shared text on the internet are for the most part free game whether we like it or not. If you're claiming that this material is not enough to train even the least powerful LLMs, then yeah, they shouldn't exist.

   • Embed this notice
     Rich Felker (dalias@hachyderm.io)'s status on Tuesday, 25-Aug-2026 04:21:50 JST Rich Felker

     in reply to

     • millennial fulcrum

     @thestrangelet @falcennial From what I can find, it seems dubious whether it should fall under the definition of "large", and it looks like they're handwaving a lot about whether they actually have rights to the corpus.

     They say "KL3M doesn't 'scrape' websites in violation of their terms of services or use policy." but this suggests to me that maybe they do scrape and just deem their scraping not to violate ToS. AFAICT this might just mean they honor robots.txt, but lack of a robots.txt directive not to scrape does not imply license to produce derivative works.

   • Embed this notice
     TheStrangelet :bc: (thestrangelet@beige.party)'s status on Tuesday, 25-Aug-2026 04:21:52 JST TheStrangelet :bc:

     in reply to

     • Rich Felker
     • millennial fulcrum

     @dalias @falcennial You are right that just because a work is public on the internet, doesn't mean it can be used in any way for any result. So scratch that. However saying it's impossible to create a model without theft is also false, here's an example where it's been done: https://www.wired.com/story/proof-you-can-train-ai-without-slurping-copyrighted-content/. You can probably shift the goalposts and say, no "general use LLM," but again there nothing inherant about the technology that FORCES theft. Yeah, it's a lot easier to create an LLM if you do steal data. It would also be a lot easier for me to be a multi-millionaire if I robbed banks on the regular.