3 ms·

There is simple way to fix this. Ban private large models trained on public data, require them to be public weights. If a company wants to train large private

There is simple way to fix this. Ban private large models trained on public data, require them to be public weights. If a company wants to train large private model, they can do it with their own data.

low_tech_punk 3y ago
Some of the public data may still have license attached, so there are still challenges in provenance, attribution, and copyright. To draw a parallel in software, we have MIT licenses that allow for-profit, private use of source code in the public. The copyleft license might be more aligned to what you are envisioning?
- mirekrusin 3y ago
  You can draw parallel with many things, ie. student who is learning on copyrighted work in a library.
whywhywhywhy 3y ago
For some reason the people complaining don’t agree with that either.
- mirekrusin 3y ago
  You can't make everybody happy at the same time.
niceice 3y ago
It's not simple since you can't ban it everywhere. China, India, et al. You can't ban it for open source either.
- s1artibartfast 3y ago
  I agree that it isnt that simple, but you could still ban it. Real world policy is rarely black and white, all or nothing. An inability to prevent every instance of something is not a reason to forgo any action Just because country X has slavery, doesnt mean the US can't ban it. There may be some consequences, and disadvantages to doing so, but that is a trade off, not a show stopper. Same thing with open source. There is lots of illegal and bootlegged content that is shared on the internet, and cant be eradicated. That doesnt mean it can all be freely bought at Walmart.
- mirekrusin 3y ago
  It doesn't matter, those states would not be better off if they did not ban it. They would be better off if you did it the other way around - ban large public models.
amenhotep 3y ago
I don't see many people suggesting this, but I also quite like this way of thinking about it. The idea that it should be illegal for models to learn from artists, or that artists have a right to extract payment from the model, doesn't make much sense to me, it's too much of a radical departure from the way we treat human learning. But it seems unfair that a company can own such a model. It's not their work, it's a codified expression of basically our entire cultural output as a species. We should all own it.
- deely3 3y ago
  > We should all own it. Maybe, just maybe better to ask artist, who should own it?
  - CaptainFever 3y ago
    Which artist? There are billions.
    - deely3 3y ago
      I think all of them? We can fix it. Lets create "for-ai-training" tag. With default value "false", so, when tag is absent you can't use text and images for ai training.
- elicksaur 3y ago
  Machine learning is a radical departure from human learning, so it seems like different rules make sense.
Workaccount2 3y ago
The simple fix is another AI that checks the output for copyright violation. The issue here is models generating copy written work verbatim. People claiming that training on copyrighted work is a violation of copyright (its not) have no legal legs to stand on. They are purposelessly muddying concepts though to make it seem like it is. However any competent judge is going to see right through this.
CrimsonRain 3y ago
If you want a large model with public weights, just do it yourself. Nobody is stopping you.
- fragmede 3y ago
  Cool. Just give me $500k for GPU time and I'll get started.
- esafak 3y ago
  At the least, we need a Creative Commons license for ML: if you train on my data, your model must be public and permit remixing. And I am in favor of more stringent licenses than that: you must pay me to use my data.
CaptainFever 3y ago
This is a great compromise, honestly. Only slightly related, but what if copy"right"-law was opt-out? You could ignore it if you want, but you aren't allowed to hold any either. I would join.
notfed 3y ago
I don't think this fixes much? People will just use open source image generators. Artists still have the same dilemma.

low_tech_punk 3y ago

Some of the public data may still have license attached, so there are still challenges in provenance, attribution, and copyright. To draw a parallel in software, we have MIT licenses that allow for-profit, private use of source code in the public. The copyleft license might be more aligned to what you are envisioning?

mirekrusin 3y ago
You can draw parallel with many things, ie. student who is learning on copyrighted work in a library.

whywhywhywhy 3y ago

For some reason the people complaining don’t agree with that either.

mirekrusin 3y ago
You can't make everybody happy at the same time.

niceice 3y ago

It's not simple since you can't ban it everywhere. China, India, et al. You can't ban it for open source either.

s1artibartfast 3y ago
I agree that it isnt that simple, but you could still ban it. Real world policy is rarely black and white, all or nothing. An inability to prevent every instance of something is not a reason to forgo any action Just because country X has slavery, doesnt mean the US can't ban it. There may be some consequences, and disadvantages to doing so, but that is a trade off, not a show stopper. Same thing with open source. There is lots of illegal and bootlegged content that is shared on the internet, and cant be eradicated. That doesnt mean it can all be freely bought at Walmart.
mirekrusin 3y ago
It doesn't matter, those states would not be better off if they did not ban it. They would be better off if you did it the other way around - ban large public models.

amenhotep 3y ago

I don't see many people suggesting this, but I also quite like this way of thinking about it. The idea that it should be illegal for models to learn from artists, or that artists have a right to extract payment from the model, doesn't make much sense to me, it's too much of a radical departure from the way we treat human learning. But it seems unfair that a company can own such a model. It's not their work, it's a codified expression of basically our entire cultural output as a species. We should all own it.

deely3 3y ago
> We should all own it. Maybe, just maybe better to ask artist, who should own it?
- CaptainFever 3y ago
  Which artist? There are billions.
  - deely3 3y ago
    I think all of them? We can fix it. Lets create "for-ai-training" tag. With default value "false", so, when tag is absent you can't use text and images for ai training.
elicksaur 3y ago
Machine learning is a radical departure from human learning, so it seems like different rules make sense.

Workaccount2 3y ago

The simple fix is another AI that checks the output for copyright violation. The issue here is models generating copy written work verbatim. People claiming that training on copyrighted work is a violation of copyright (its not) have no legal legs to stand on. They are purposelessly muddying concepts though to make it seem like it is. However any competent judge is going to see right through this.

CrimsonRain 3y ago

If you want a large model with public weights, just do it yourself. Nobody is stopping you.

fragmede 3y ago
Cool. Just give me $500k for GPU time and I'll get started.
esafak 3y ago
At the least, we need a Creative Commons license for ML: if you train on my data, your model must be public and permit remixing. And I am in favor of more stringent licenses than that: you must pay me to use my data.

CaptainFever 3y ago

This is a great compromise, honestly. Only slightly related, but what if copy"right"-law was opt-out? You could ignore it if you want, but you aren't allowed to hold any either. I would join.

notfed 3y ago

I don't think this fixes much? People will just use open source image generators. Artists still have the same dilemma.