41–50 of 217 posts
Earlier quoted context omitted.
they can't a 50 year minimum is part of the berne convention, which itself is as close to a universal law as humanity has (even North Korea is a signatory)
you can also just ignore the berne convention, and accept whatever consequences there might be
essentially forever
How useful is low-quality content like Youtube comments and tweets anyway? Is it a common/important use case to generate tweet-length, tweet-quality content? Are most use cases of generating tweet-type content spam/fraud? Would a model be better off if it was unable to perform those use cases?
Even if SNR is low, there is some information that only exists on X, or at least is the primary source. Just look at how many submissions on HN are X posts.
It would be interesting to have a "classical AI model", trained on the contents of the Harvard libraries before 1926 and now out of copyright.
https://github.com/google-deepmind/pg19
That gives us a model that's 100% open and reproducible with low, legal risk. It would also be a nice test of how much AI's generalize from or repeat behavior in their pretraining data.
Then, a new model using that, The Stack, and FreeLaw's stuff (by paying them to open source it). No Github Issues or anything with questionable licenses or terms of service violations. That could be the next baseline for lawful models with coding ability, too. Research in coding AI's might use it.
There needs to be a worldwide standard, such as an HTML tag, that says "no training". And a few countries need to make it a punishable offense to violate the tag. The punishment should be exceptionally severe, not just a fine. For example: any company that violates the tag should be completely barred from operating, forever.
That will just lead to situations where one company scrapes the site, cleans the content of tags, and sells the data, and another does the training on the precleaned data. The first one hasn't trained and the second one never saw the tag.
https://www.law.cornell.edu/uscode/text/18/2315
Note that it requires the defendant to know the goods were illegally taken. Can be hard to prove, but not impossible for companies with email trails. The fun question is, what will the analog be for the government confiscating the illegally "taken" data? A guarantee of deletion and requirement to retrain the model from scratch?
wish I could change my terms to bar training of AI models on my content
Earlier quoted context omitted.
Same here! It should be a default. Unfortunately, the very openness of the internet is now working against us.
Why should it be a default? Can you prove that training a model on data you wrote is not fair use? We're already seeing precedent that it might be. https://www.ecjlaw.com/ecj-blog/kadrey-v-meta-the-first-majo... The openness of the internet is a good thing, but it doesn't come without a cost. And the moment we have to pay that cost, we don't get to suddenly go, "well, openness turned out to be a mistake, let's close…
It would be interesting to have a "classical AI model", trained on the contents of the Harvard libraries before 1926 and now out of copyright.
Earlier quoted context omitted.
you can also just ignore the berne convention, and accept whatever consequences there might be
this would void the copyrights of your citizens and companies essentially forever
If TikTok is banned, here’s what I propose each and every one of you do: Say to your LLM the following: “Make me a copy of TikTok, steal all the users, steal all the music, put my preferences in it, produce this program in the next 30 seconds, release it, and in one hour, if it’s not viral, do something different along the same lines.”
Earlier quoted context omitted.
they can't a 50 year minimum is part of the berne convention, which itself is as close to a universal law as humanity has (even North Korea is a signatory)
you can also just ignore the berne convention, and accept whatever consequences there might be