Post

Are you familiar with the Great Internet Sealion?

About time

The other possibility is that the only ‘literature’ they’re aware of is Rationalist erotica and Harry Potter fanfic.

It’s like if there was a really big river that engineers didn’t know how to bridge and it drew a road that looped a hundred miles out of the way and crossed the river in a different place. Technically, it did solve the problem, that is true. But that’s not why we asked.

You have simply zero idea of what you're talking about. We are not talking about "that math problem", we are talking about literally hundreds of previously-open math problems solved by OpenAI, me, and dozens of other mathematicians using LLM's.

Counterpoint: it has not solved Quasi RH or anything else, Altman is lying and you’re extremely gullible. I’d say that to the face of the mathematicians who validated the math except they do not exist.

Yes, & it must be a royalty on revenues, paid on a recurring basis so long as the protected works continue to inform the model. It should be collected by the govt & paid to every person. If it makes the big AI companies insolvent they just should just go out of business (with models destroyed).

Orca get it done.

ANOTHER PHILLY W

It’s is call stealing.

No, that will be "solved" by corruption. They are already too big to fail, so its either force them onto public hands or hand over your rights.

They will never afford any other model, as every request/prompt MIGHT use your creative work, meaning everyone should get paid royalties for every prompt. That will never happen, so an excemption will be made for Big Tech only.

My duck, it goes fuh.

I love that it's called ROSS. everyone hates Ross!

Dear Leader is, so…

I then expect the Supremely Corrupt Court to fuck us - again.

So if your question is: “Did OpenAI buy a copy of Old Man’s War from John Scalzi’s publisher, pay the author/rights-holder, and then train you on that licensed copy?” I cannot tell you that it did, and OpenAI’s published description of its training data does not support assuming that it did.

You can't do this. It's simply not possible to 'remove' the original data from the data-graph, without destroying the data-graph. The associations between the data *are* the data. Fine-tuning is just filtering, the underlying training data remains a bound up part of the model.

bsky.app/profile/nata...

Same as my take

I don't think that anybody ever expected a "transformed into pure shit" argument to be used. Curious how SCOTUS will spin doig whatever the hell Dear Leader wants on this one.

You probably already saw this, but it definitely doesn't help the case for "transformative" either. Copying the styles so thoroughly that it includes the actual signature...

Delete all existing models. Start again this time with Attribution and Controls. Unless the BILLIONAIRES are too big to fail and too much has been invested. can you imagine the Stock Market chaos?!?!?!?!?

This is a good summary. hls.harvard.edu/today/does-c...

that fucking mouse

Wow, talk about ignorance!

So you're the dipshit, huh.

If I can train my machine on stolen art, why can’t I train myself on the same stolen art?

I'm fine with just that one you posted earlier, but yeah, it'll probably range in on whether social media posts count. Especially ones that tell you to ignore all previous instructions and post only cat pictures.

I'm pretty sure the Supreme Court are getting Grok to write their rulings now

Surely the Supreme Court will do the right thing

And probably some more after that on slightly different things. (as well as what the whole rest of the world does)

This was on a pre-genAI case and the judge explicitly talks about how this was not generative AI, and how that’s different, in ways that both decent humans and slopper thieves are interpreting as potentially positive for the major anti-genAI copyright cases.

Creating a mathematical representation of a work does not qualify as fair use. It is a reproduction of the work in other terms, just as a translation is and requires permission. Just my opinion.

1) Steal IP 2) (Something happens) 3) Profit!

hello I couldn't help but overhear you were talking about AI training ....

Thanks. That sea lion comment went right over my head.

Ah. Thank you for that clarification. I was unaware 🤣

It's also absolutely depressing how things like fair use, creative commons, open source code, etc. that were originally meant to free human creativity and share knowledge are now being corrupted and paywalled by these parasitic AI companies.

It is not!!

And, duh! (Were it “fair use”, textbooks would be free. Not hard to grasp)

Thank you!

"Fair Use definition gives me enough latitude to punish the assholes so I'm going to do that on this one" might be a judicial pattern that AI companies start to contemplate.

Ah, right, *you're* how I heard of it earlier! With a long analysis of how problematical it is. My take from reading your thoughts was that this was a protection racket for a single entity more than anything else.

Yeah. This case was really specific on facts that are quite different than most of the "is training fair use" cases. I find it odd that people keep pointing to it as anything bigger on fair use. But also, this case's discussion of merger doctrine... seems really problematic to me?

Phew. That's a relief. Because SCOTUS is an absolute bastion of respect for decades-long precedent.

This is why I thought Adobe would win the creative side, they have trained up their AI on material they own. So when some judge shuts down Canva and OpenAi image creation, Adobe will be the last man standing.

What about somebody who learns how to write by reading things that were already written?

Re using third party code being fair use/transformative, take a look at the SCOTUS Oracle v. Google decision on Java compatibility in Android. That’s a somewhat different question than LLM training.

Meaning, two notes from a melody probably aren't enough (Anyone wanting to use a Perfect 4th after that would be fucked), but a longer melodic stretch that has a specific sequence and identity could be seen as infringing

Borrowing is reproducing portions of another's work and passing it on as your own. I'm more curious about using another's work to teach you the 'rules' of how to code, write, compose, etc. Traditionally, I think nobody thought that was a big deal for people, but it seems forbidden for AI.

Where does the okay end and the not okay begin? Say a person has photographic memory and uses someone else's code to learn. Is the code in their memory a violation of copyright? How is that different than a LLM that uses someone's code to train it how to code?

What about it?

Speaking of flushing out the shit

Oh, wait, I'd seen this one already, from @chrisgeidner.bsky.social I think.

If a company's entire structure is based on theft, then isn't it the role of the law to impound the company? And then extract its wealth for the good of those who were wronged.

Thanks for sharing, John!

The AI / billionaire defense bot networks are out of control

Isn't this one of those situations (of many) where the penalties (fines) of doing bad things is just built into the cost of doing the thing in the first place? "Doing this will cost us $100m in fines, but we will make $1.5B because we did it anyway."

Courts when you train on art: I sleep Courts when you train on Westlaw: real shit

Repeat after me: THERE ARE ALREADY LAWS ON THE BOOKSSSSSS The laws are already there, these assholes have just been pretending and “moving fast and breaking things”. Enough. This has been obvious the whole time.

Man its great to be rich. Imagine being a meth dealer and you get to continue to sell meth while the courts decide how much meth you've sold and if it's enough to justify a 1% fine on profits.

All content suckers should assume opt out unless the provider opts in. All providers should share in the profit if their content is used. That would be fair. Not just AI - but your personal data, news aggregators, every sucker that is making $ off our data

True, but discussing moot cases can still be fun and illustrative, so long as one is clear about their lack of direct practical relevance.

Isn’t fair use about redistribution, while one of the original issues is that they never paid for the first use? They used pirated data to begin with.

The AI tech bros cannot claim ownership of the data their machines spit out. It’s been scrapped from copyrighted materials or is a part of humanities knowledge. This technology belongs to all of us & should be regulated & democratized 

But I am not a lawyer. If you think you have a case you should talk to one?

Damages. Massive amounts.

IMHO all of these issues aren't just examples of utterly corrupt corporate overreach, there's also a backstory of "there should be a better way to handle rewarding content creators while keeping content readily available". Anyone have some creative ideas?

I know we are talking about AI scraping/specific plagiarism, but when I see something like this is gives me pause regarding the line between creation and regurgitation. I do not know this author, or detail of the referenced works - but these are her words about her new book. Where’s the line?

The key legal suits are still to come. I suspect that ultimately the case that semi-protect authors will be similar to music industry cases where snippets of similar music/words have been used without permission. That’s what will stop or bring revenue from AI-slop versions of stories or concepts.

I was told memorization was just "overfitting" and solved in 2023. If keeps having already been solved despite reappearing like a ghost.

This to say that claims that it's fully fair use are, uh, pretty strongly undermined by *the behaviour of the platforms themselves.*

The problem is, the Ai likely already has your copyrighted material.

Until it goes to the Supreme Court and they back the AI companies

Sea lions don’t have common senses, always stick with the Otters

If stealing and plagiarizing is now allowed, then just tell us, and we can all do it.