3 ms·

I can get that. There's never a path forward that doesn't leave something on the table. I of course want artistic freedom + technology to advance alongside each

I can get that. There's never a path forward that doesn't leave something on the table. I of course want artistic freedom + technology to advance alongside each other. But we're also talking about MAKING generative AI models. Not USING generative AI models. Making a model is so far outside of the scope of a single artist it's almost comical. We're talking about a task that, given no bounds, would include the subtask "ingest all human creative output". That's hard to do so therefore it's restricted to the realm of well funded companies. But even as well off and talented as these companies are, they did very little of the actual work. The total scale of effort that goes into these big lists of coefficients, is beyond human comprehension at this point. We've literally NEVER had to contend with tasks of this magnitude before. Many thousands of centuries of human life hours distilled. Copyright as it's legally inscribed precisely, is ill-prepared to achieve it's end goals in this situation. But considering: 1) that's kind of cool we can do this 2) it makes interesting things that could better people's lives 3) we still think that humans should be able to make art (ai or not) and not die of starvation ideally ... we need to figure out what to do. Maybe it's not copyright! That was created to solve a specific problem. Now we have new problems. I just think people latch too hard onto the legalese as if it's some magic property of the universe, and NOT a set of systems humans made, for very human goals. I don't want a system where we just say "well look at copyright, it says this is fine, so we will just use everything for free". It's missing the point. I agree with you that it's asking the hard question of "which artists do we protect". That, I'm not feeling like I have a good answer for right now. I definitely heavily lean towards the talented artists who've honed their craft. That's just more interesting. But I wish more people we're thinking about it, in those terms.

visarga 3y ago
> We've literally NEVER had to contend with tasks of this magnitude before. Many thousands of centuries of human life hours distilled. And one poem or a page of text is but a drop in this ocean, an infinitesimal contribution. The model is much smaller than the training set. For example Stable Diffusion packs 5B images in 5GB, that's a compression factor of 100,000:1, about 1 byte per training image, not even a pixel. Taking so little from everyone, I am amazed how strong a reaction it creates. The final model doesn't steal almost anything, it doesn't rely on any training example too much, just a compressed abstraction of the training data. You have to overfit it and also prime it to make it regurgitate some training examples, and most attempts fail, resulting in novel outputs.

visarga 3y ago

> We've literally NEVER had to contend with tasks of this magnitude before. Many thousands of centuries of human life hours distilled. And one poem or a page of text is but a drop in this ocean, an infinitesimal contribution. The model is much smaller than the training set. For example Stable Diffusion packs 5B images in 5GB, that's a compression factor of 100,000:1, about 1 byte per training image, not even a pixel. Taking so little from everyone, I am amazed how strong a reaction it creates. The final model doesn't steal almost anything, it doesn't rely on any training example too much, just a compressed abstraction of the training data. You have to overfit it and also prime it to make it regurgitate some training examples, and most attempts fail, resulting in novel outputs.