Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Alignment whack-a-mole: Finetuning activates recall of copyrighted books in LLMs

github.com

Post reply on HN
Print view View original

171–180 of 184 posts

Re: Alignment whack-a-mole: Finetuning activates recall of copyrighted books in LLMs

#171

Post by mock-possum »

palmotea wrote: post #63
userbinator wrote: post #46

Earlier quoted context omitted.

IMHO giving many details in the prompt and asking the model to "fill in the blanks" feels a little like cheating in the same way as embedding the dictionary in the decompression program. But it will certainly make the Imaginary Property lawyers squirm.

It's not cheating, it seems like a technique to defeat obfuscation to show the content is there in a complete or near-complete form, which proves it was copied.

What good is a technique to defeat obfuscation, if it requires you to already have the information deobfuscated?

Re: Alignment whack-a-mole: Finetuning activates recall of copyrighted books in LLMs

#172

Post by palmotea »

mock-possum wrote: post #171
palmotea wrote: post #63

Earlier quoted context omitted.

It's not cheating, it seems like a technique to defeat obfuscation to show the content is there in a complete or near-complete form, which proves it was copied.

What good is a technique to defeat obfuscation, if it requires you to already have the information deobfuscated?

> What good is a technique to defeat obfuscation, if it requires you to already have the information deobfuscated?

You misunderstand the goal, which is to prove that an unlicensed copy is stored in the system.

It's like a scenario where Alice sent another Bob some highly illegal data that's a crime to possess, and encrypted it with Bob's public key. The police got a warrant and have access to Bob's email but not his secret key. They can prove he has the data (and convict him) if they have a copy, encrypt it again with his public key, and compare the files and find them to be identical. Could they have gotten the unencrypted from Bob's email? No, but they didn't have to.

Re: Alignment whack-a-mole: Finetuning activates recall of copyrighted books in LLMs

#173

Post by Leynos »

lelanthran wrote: post #127
TFNA wrote: post #11

I’m a researcher who for years has been scanning my library’s holdings on my particular discipline for my own use, but also uploading the books to the shadow libraries for everyone else’s benefit. The revelation that LLMs are training on the shadow libraries has made me put a lot more effort into ensuring my scans are well-OCRed. The idea that I could eventually ask ChatGPT or whatever about obscure things in my fiel…

> The idea that I could eventually ask ChatGPT or whatever about obscure things in my field, and get useful output (of the "trust but verify" sort), is exciting. That's your idea, not the one they are going with. Their idea is that you pay a fee to access any information that was freely available. Your idea is tearing down of fences, their idea is gatekeeping. The two ideas are incompatible.

The library doesn't cease to exist

Re: Alignment whack-a-mole: Finetuning activates recall of copyrighted books in LLMs

#174

Post by Leynos »

soco wrote: post #134
BugsJustFindMe wrote: post #41

Earlier quoted context omitted.

Library funding is a political stance that has only imaginary connection to whether people pay to ask things of ChatGPT. People can pay to talk to an AI and also government can fund libraries.

The government can then soon "optimize" and fund exactly one library.

If you vote for people who think like this, you have to face the consequences of your actions.

Re: Alignment whack-a-mole: Finetuning activates recall of copyrighted books in LLMs

#175

Post by red75prime »

palmotea wrote: post #172
mock-possum wrote: post #171

Earlier quoted context omitted.

What good is a technique to defeat obfuscation, if it requires you to already have the information deobfuscated?

> What good is a technique to defeat obfuscation, if it requires you to already have the information deobfuscated? You misunderstand the goal, which is to prove that an unlicensed copy is stored in the system. It's like a scenario where Alice sent another Bob some highly illegal data that's a crime to possess, and encrypted it with Bob's public key. The police got a warrant and have access to Bob's email but not his…

No need for analogies. They use GPT-4o to generate summaries from verbatim quotes. Then those summaries (which are comparable in length to the verbatim quotes) are used to generate hundreds of passages in the style of an author and they use verbatim quotes to find exact matches. They don't control whether this procedure introduces an information leak that makes exact generation more likely.

Re: Alignment whack-a-mole: Finetuning activates recall of copyrighted books in LLMs

#176

Post by Gander5739 »

fragmede wrote: post #97
GaryBluto wrote: post #65

Earlier quoted context omitted.

> How do you deal with the ethics of this personally, stealing to make it easier for AI to steal so AI gets better? By quoting your comment in my reply, have I "stolen" your comment?

By reading this comment you have entered into a legal contract, by which you owe me $5. Failure to pay will be reported to the Internet police.

https://xkcd.com/501/

Re: Alignment whack-a-mole: Finetuning activates recall of copyrighted books in LLMs

#177

Post by lofaszvanitt »

xerox13ster wrote: post #161
maplethorpe wrote: post #149

Earlier quoted context omitted.

What about something you've made for fun but haven't made any money from? Should someone else be allowed to sell and profit from your work? I'm not expecting to be "passively paid" for my hobbies. But I'm expecting that someone won't steal and profit from the things I make. Why would that be fair?

Say my hobby is statue making. I design and create a concrete statue that I failed to sell. Whether that be because I did not try or because I could not find a buyer, I could not sell it. So I took it, and I put it in my pile of completed works: a pile of crumbling statute rubble by the roadside. In the digital case, maybe it was posted online and the pile is a timeline or portfolio. Someone in a pick up truck drives…

Are you high or a paid shill?

Re: Alignment whack-a-mole: Finetuning activates recall of copyrighted books in LLMs

#178

Post by lofaszvanitt »

xerox13ster wrote: post #162
fibonacci_man wrote: post #157

Earlier quoted context omitted.

So no authors, directors, or any other creative work that can be stolen & duplicated? Why don’t we get rid of patent laws too while we’re at it?

Unironically, let’s get rid of patent laws while we’re at it. The advancement of technology would take off if we did not have patent trolls telling us what you could, and could not use understand and improve. Just imagine what Palworld could be if it didn’t have to spend the last year two years, however long spending all of their budget fighting a patent case against the biggest fucking gaming company in the world in…

Palworld ripped off another company's hard work. The IP holders poured insane amounts of time and money to market and make their work a success story. Now comes a bunch of snotty kids and try to make money off it. Nah, go fuck off. Use your head and come up with something original.

AI already gives IQ to the braindead masses. And now people want to abolish copyright, so the totally unworthy, useless nobodies could leech other's hard work and profit from it. Get a grip.

Maybe do not make games if you have zero ideas yourself. Go and mow the lawn if that's all you are capable of.

Re: Alignment whack-a-mole: Finetuning activates recall of copyrighted books in LLMs

#179

Post by lofaszvanitt »

SomaticPirate wrote: post #62
ralph84 wrote: post #42

Earlier quoted context omitted.

OpenAI's valuation is more than basically all traditional media companies combined. Nvidia could buy the NYTimes with a month's worth of profits. The top 8 companies in the S&P 500 all benefit more from LLMs being successful than strict copyright enforcement. Congress has very broad power over copyright law. If a suit is successful there is a lot of money and power to be deployed to change copyright law.

Exactly. So just buy it. They have the money or does Sam need a moonbase to complete his villain arc. Any of these AI companies could come out and start paying creators a licensing fee. Instead of being forced to pay damages which is their current approach

But they don't want to. That's their business model.

Re: Alignment whack-a-mole: Finetuning activates recall of copyrighted books in LLMs

#180

Post by xerox13ster »

lofaszvanitt wrote: post #178
xerox13ster wrote: post #162

Earlier quoted context omitted.

Unironically, let’s get rid of patent laws while we’re at it. The advancement of technology would take off if we did not have patent trolls telling us what you could, and could not use understand and improve. Just imagine what Palworld could be if it didn’t have to spend the last year two years, however long spending all of their budget fighting a patent case against the biggest fucking gaming company in the world in…

Palworld ripped off another company's hard work. The IP holders poured insane amounts of time and money to market and make their work a success story. Now comes a bunch of snotty kids and try to make money off it. Nah, go fuck off. Use your head and come up with something original. AI already gives IQ to the braindead masses. And now people want to abolish copyright, so the totally unworthy, useless nobodies could le…

May you ever be the defendant of a patent troll's lawsuit on tech you believed was original.
Post reply on HN