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

Stable Audio Open

stability.ai

Post reply on HN
Print view View original

51–60 of 137 posts

Re: Stable Audio Open

#51

Post by lolinder »

avarun wrote: post #30
Uehreka wrote: post #25

Earlier quoted context omitted.

The environmental impact. And yes, I know, 0.5%, but my issue was always that if PoW currencies went from being a niche subculture to a point where it was used for everyday exchange (many people were arguing that this would and should happen) that 0.5% would surely go up by a great deal. To a point where crypto had to clear a super high bar of usefulness to counterbalance the harm it would do. To be fair, AI training…

There is no "environmental impact". Environmental impact comes from energy production, not energy usage. It's incoherent to argue others should tamper down their energy usage because most folks producing energy aren't doing it in an ethical way.

> It's incoherent to argue others should tamper down their energy usage because most folks producing energy aren't doing it in an ethical way.

There's a general consensus that paying someone else to do your dirty work doesn't free you of the moral (or, usually, legal) culpability for the damage done. If you knowingly direct your money towards unethical providers, you are directly increasing the demand for unethical behavior.

(That's assuming that the producers themselves are responsible for the ethics. If a producer is doing its best to convert to clean energy as fast as possible, they may be entirely in the clear but POW would still be unethical. In that scenario POW is placing strain on the limited clean energy supplies, forcing the producer to use more fossil fuels than they'd otherwise need to.)

Re: Stable Audio Open

#52

Post by jncfhnb »

avarun wrote: post #30
Uehreka wrote: post #25

Earlier quoted context omitted.

The environmental impact. And yes, I know, 0.5%, but my issue was always that if PoW currencies went from being a niche subculture to a point where it was used for everyday exchange (many people were arguing that this would and should happen) that 0.5% would surely go up by a great deal. To a point where crypto had to clear a super high bar of usefulness to counterbalance the harm it would do. To be fair, AI training…

There is no "environmental impact". Environmental impact comes from energy production, not energy usage. It's incoherent to argue others should tamper down their energy usage because most folks producing energy aren't doing it in an ethical way.

Officer I merely stabbed the man. What he died from was blood loss.

Re: Stable Audio Open

#53

Post by SahAssar »

Workaccount2 wrote: post #37
JohnKemeny wrote: post #29

Earlier quoted context omitted.

I think you should read the case material for NY Times v OpenAI and Microsoft. It literally says that within ChatGPT is stored, verbatim, large archives of NY Times articles and that they were able to retrieve them through their API.

..which makes no sense. It is either an argument of ignorance or of purposeful deceit. There is no coherent data corpus (compressed or not) in ChatGPT. What is stored are weights that create a string of tokens that can recreate excerpts data that it was trained on, with some imperfect level of accuracy. Which I agree is problematic, and OpenAI doesn't have the right to disseminate that. But that doesn't mean OpenAI d…

> Content creators are doing a purposeful slight of hand to confabulate "outputting copyrighted data" with "training on copyrighted data".

I don't think so, I think it's usually argued as two different things.

The "training on copyrighted data" argument is usually that we never licensed this work for this sort of use and it is different enough from previously licensed uses that it should be treated differently.

The "outputting copyrighted data" argument is somewhat like your output is so similar as to constitute a (at least) partial copy.

Another argument is that licensed data is whitewashed by being run through a model. So you could have GPL licensed code that is open source run through a model and then output exactly the same but because it has been outputted by the model it is considered "cleaned" from the GPL restrictions. Clearly this output should still be GPL:ed.

> It's not illegal for me to read an NYT article and write my own summary of the article's contents on my blog. This has been true forever and has forever been a staple in new content creation.

What if I compress the NYT article with gzip? What if I build a LLM model that always replies with the full article within 99% accuracy? Where is the line?

This is not a technical issue, we need to decide on this just like we did with copyright, trademarks, etc. Regardless of what you think this is not a non-issue and we cant use the same rules as we did up until now unless we treat all ML systems as either duplication or humans and neither seems to solve the issues.

Re: Stable Audio Open

#54

Post by tomcam »

Workaccount2 wrote: post #24
Uehreka wrote: post #11

> The new model was trained on audio data from FreeSound and the Free Music Archive. This allowed us to create an open audio model while respecting creator rights. This feels like the “Ethereum merge moment” for AI art. Now that there exists a prominent example with the big ethical obstacle (Proof of Work in the case of Ethereum, nonconsensual data-gathering in the case of generative AI) removed, we can actually have…

The idea that AI trained on artist created content is theft is kind of ridiculous anyway. Transformers aren't large archives of data with needles and thread to sew together pieces. The whole argument is meant to stifle an existential threat, not to halt some illegal transgression. If they cared about the latter a simple copyright filter on the output of the models would be all that's needed.

Yet before “safeguards” were added a prompt could say “in the style of Studio Ghibli” and you could get exactly that.

Would it be possible if Studio Ghibli images had not been used in the training?

Re: Stable Audio Open

#55

Post by rfoo »

pa7x1 wrote: post #36
rfoo wrote: post #14

Earlier quoted context omitted.

Why is Proof of Work less ethical than Proof of Rich a.k.a. rich being gradually more rich without doing anything? Not saying PoW is safer (it's not), but less ethical is pretty a bold claim.

How do the rich become gradually richer under PoS? I'm flabbergasted by the level of math education. Assume we have 2 validators in the network; the first one owns 90% of the network, the second one owns 10%. Lets call them Whale and Shrimpy, respectively. To make the numbers round let's assume total circulating supply of ETH is 100 initially and that the yield resulting from being a validator is 10% per year. After…

Sure, friends also won't let friends skip the fact that circulating supply of ETH is now decreasing instead of increasing.

Also, only ~30% tokens are staked. The 30% who chose to stake essentially tax the other 70% in use. Each of the validator do the same amount of work (ok, strictly speaking you get to do more when you have more ETH staked, but being a validator is cheap and does not cost significantly more energy even if you are being selected more frequently because running one proposal is too cheap, that's the whole environmental point, right?) except what they receive is proportioned to how much they stake.

I hate being mean, but sorry, remembering to check one's assumption is a habit I gained after elementary school, so maybe that's too hard for you.

Re: Stable Audio Open

#56

Post by semi »

tomcam wrote: post #54
Workaccount2 wrote: post #24

Earlier quoted context omitted.

The idea that AI trained on artist created content is theft is kind of ridiculous anyway. Transformers aren't large archives of data with needles and thread to sew together pieces. The whole argument is meant to stifle an existential threat, not to halt some illegal transgression. If they cared about the latter a simple copyright filter on the output of the models would be all that's needed.

Yet before “safeguards” were added a prompt could say “in the style of Studio Ghibli” and you could get exactly that. Would it be possible if Studio Ghibli images had not been used in the training?

if it was trained on a sufficient amount of fan art made in the studio Ghibli style and tagged as such, yes.

otherwise those would just be unknown words, same as asking an artist to do that without any examples.

though I am curious how performance would differ between training on only actual studio Ghibli art, only fan art, or a mix. Maybe the fan art could convey what we expect 'studio Ghibli style' to be even more, whereas actual art from them could have other similarities that that tag conveys.

Re: Stable Audio Open

#57

Post by jrm4 »

Workaccount2 wrote: post #24
Uehreka wrote: post #11

> The new model was trained on audio data from FreeSound and the Free Music Archive. This allowed us to create an open audio model while respecting creator rights. This feels like the “Ethereum merge moment” for AI art. Now that there exists a prominent example with the big ethical obstacle (Proof of Work in the case of Ethereum, nonconsensual data-gathering in the case of generative AI) removed, we can actually have…

The idea that AI trained on artist created content is theft is kind of ridiculous anyway. Transformers aren't large archives of data with needles and thread to sew together pieces. The whole argument is meant to stifle an existential threat, not to halt some illegal transgression. If they cared about the latter a simple copyright filter on the output of the models would be all that's needed.

I fail to see how the argument is ridiculous; and I'll bet that a jury would find the idea that "there is a copy inside" at least reasonable, especially if you start with the premise that "the machine is not a human being."

What you're left with is a machine that produces "things that strongly resemble the original, that would not have been produced, had you not fed the original into the machine."

The fact that there's no "exact copy inside" the machine seems a lot like splitting hairs; like saying "Well, there's no paper inside the hard drive so the essence of what is copyable in a book can't be in it"

Re: Stable Audio Open

#58

Post by kimixa »

Workaccount2 wrote: post #24
Uehreka wrote: post #11

> The new model was trained on audio data from FreeSound and the Free Music Archive. This allowed us to create an open audio model while respecting creator rights. This feels like the “Ethereum merge moment” for AI art. Now that there exists a prominent example with the big ethical obstacle (Proof of Work in the case of Ethereum, nonconsensual data-gathering in the case of generative AI) removed, we can actually have…

The idea that AI trained on artist created content is theft is kind of ridiculous anyway. Transformers aren't large archives of data with needles and thread to sew together pieces. The whole argument is meant to stifle an existential threat, not to halt some illegal transgression. If they cared about the latter a simple copyright filter on the output of the models would be all that's needed.

I think the definition between "Lossy Compression" and "Trained AI" is... vague according to the current legal definitions. Or even "lossless" in some cases - as shown by people being able to get written articles output verbatim.

While the extremes are obvious, there's a big stretch of gray in the middle. A similar issue occurs in non-AI art, the difference between inspiration and tracing/copying isn't well defined either, but the current method of dealing with that (being on a case-by-case basis and a human judging the difference) clearly cannot scale to the level that many people intend to use these tools.

Re: Stable Audio Open

#59

Post by pa7x1 »

rfoo wrote: post #55
pa7x1 wrote: post #36

Earlier quoted context omitted.

How do the rich become gradually richer under PoS? I'm flabbergasted by the level of math education. Assume we have 2 validators in the network; the first one owns 90% of the network, the second one owns 10%. Lets call them Whale and Shrimpy, respectively. To make the numbers round let's assume total circulating supply of ETH is 100 initially and that the yield resulting from being a validator is 10% per year. After…

Sure, friends also won't let friends skip the fact that circulating supply of ETH is now decreasing instead of increasing. Also, only ~30% tokens are staked. The 30% who chose to stake essentially tax the other 70% in use. Each of the validator do the same amount of work (ok, strictly speaking you get to do more when you have more ETH staked, but being a validator is cheap and does not cost significantly more energy…

> Sure, friends also won't let friends skip the fact that circulating supply of ETH is now decreasing instead of increasing.

This changes absolutely nothing of the calculation. Furthermore, the change in circulating supply last year was of 0.07%.

> Also, only ~30% tokens are staked.

Correct.

> The 30% who chose to stake essentially tax the other 70% in use.

There is something called opportunity cost. With the existence of liquid staking derivatives the choice to stake or not is one of opportunity cost. Plenty of people may consider the return observed by staking insufficient given the opportunity cost and additional risks. Participating in staking is fully permissionless, stakers are not taxing non-stakers. They are being remunerated for their work.

> Each of the validator do exactly same amount of work (that's the point, right) except what they receive is proportioned to how much they stake.

Incorrect. A staker does proportionate amount of work to its stake. That's why it gets paid more. A staker gets paid for fulfilling its duties as defined in the protocol (attesting, proposing blocks, participating in sync committees). For each of those things there are some rewards and some punishments in case you fail to fulfill them. If a staker has more validators running you simply fulfill more of those duties more often, hence your reward scales linearly with number of validators.

Re: Stable Audio Open

#60

Post by cthalupa »

bckr wrote: post #48
Workaccount2 wrote: post #37

Earlier quoted context omitted.

..which makes no sense. It is either an argument of ignorance or of purposeful deceit. There is no coherent data corpus (compressed or not) in ChatGPT. What is stored are weights that create a string of tokens that can recreate excerpts data that it was trained on, with some imperfect level of accuracy. Which I agree is problematic, and OpenAI doesn't have the right to disseminate that. But that doesn't mean OpenAI d…

> There is no coherent data corpus (compressed or not) in ChatGPT. I disagree. If you can get the model to output an article verbatim, then that article is stored in that model. Just because it’s not stored in the same format is meaningless. It’s the same content regardless of whether it’s stored as plaintext, compressed text, PDF, png, or weights in a model. Just because you need an algorithm such as a specialized p…

> If you can get the model to output an article verbatim, then that article is stored in that model.

You can't get it to do that, though.[1]

The NYT vs OpenAI case, if anything, shows that even with significant effort trying to get a model to regurgitate specific work, it cannot do it. They found articles it had overfit on due to snippets being reposted elsewhere across the internet, and they could only get it to output those snippets, and not in correct order. The NYT, knowing the correct order, re-arranged them to fit the ordering in the article.

Even doing this, they were only able to get a hundred or so words out of the 15k+ word articles.

No one who knows anything about these models disagrees that overfitting can cause this sort of behavior, but the overwhelming majority of the data in these models is not overfit and they take a lot of care to resolve the issue - overfitting isn't desirable for general purpose model performance even if you don't give a shit about copyright laws at all.

People liken it to compression, like the GP mentioned, and in some ways, it really is. But in the most real sense, even with the incredibly efficient "compression" the models do, there's simply no way for them to actually store all this training data people seem to think is hidden in there, if you just prompt it the right way. The reality is only the tiniest fraction of overfit data can be recovered this way. That doesn't mean that the overfit parts can't be copyright infringing, but that's a very separate argument than the general idea that these are constantly putting out a deluge of copyrighted material.

(None of this goes for toy models with tiny datasets, people intentionally training models to overfit on data, etc. but instead the "big" models like GPT, Claude, Llama, etc.)

1. https://fingfx.thomsonreuters.com/gfx/legaldocs/byvrkxbmgpe/...

Post reply on HN