I think you underestimate the sheer volume of data + conclusions the brain ingests and processes on a daily basis, primarily through unconscious experience.

While our brains are more complex than the networks, this has never been in dispute.

The quantity of experiences needed to train GPT-3, however, is many more than we are capable of experiencing in a lifetime.

Also, about volume of data processed by the brain: https://gwern.net/Differences

But I think you meant GPT-3 has seen many books during training, not during inference. You should know that training on millions of books is not the only way GPT-3 learns. It is just the foundation of its knowledge.

GPT-3 learns "in-context", that means it can learn a new word or a new task at first sight. It just needs a description or a few examples. This is the most powerful feature of GPT-3 - in-context learning. And when it comes to ICL, it is much like humans - only sees a few examples, not millions of books.

> “Do you ask for permission when you train your mind on copyrighted books?”

The nature of ICL is that it happens at prediction time. So GPT-3 would have to explicitly be instructed to learn a specific skill. Should it reject instructions if they are sourced from copyrighted books?

I’m not a lawyer, but to me it seems within the realm of possibility that a U.S. court eventually finds strongly in favor of the copyright holders, the Supreme Court agrees (because Big Tech has so few friends left), and OpenAI will be required to destroy the GPT-3 model and all copies of the training data because they can’t filter out copyrighted works.

Just because you can find one way that GPT might be slow doesn't invalidate the point that its training does use massive amounts of data.

I am not a lawyer, so the following is only my opinion.

Scale matters, but so does the legality of the thing that scales.

Reading two dozen books by other authors, or studying hundreds of artworks, or visiting the museum of awesome statues every week, in order to get inspired for ones own novel/painting/scuplture, isn't illegal.

So a lawsuit will have a really hard time argueing that it somehow is a problem if its two dozen billion books/paintings/sculptures. Because, such a lawsuit would suddely need to explain why the smaller scale is also problematic, only less so. And given that this is basically how art worked ever since the first human had the idea to paint pictures on a cave wall, that's a hard sell.

Humans either learn art by being natural art geniuses, or by receiving instruction and learning through an iterative process (where, again, they might create thousands of art works, but nowhere near the scale here), which is very different.

2. An AI has a training set of every image in the world and produces an entirely unique work.

Which do you have more of a problem with?

Copyright law serves the purpose of peoples works being protected from unauthorized parties making copies of their works, and profiting off them.

It doesn't protect from technology making the production of new works cheaper, faster, more efficient. An artist using photoshop can be many times faster than one using oil and canvas.

I think the objectionable thing about the second is that the AI knows everyone's styles, and so can use them in creating something new. Even if the AI is restricted to not be able to paint an image in a certain artist style (as the new version of stable diffusion is, for instance) and the art is unique, I think part of the problem is that the AI is still (presumably) leaning on the collective styles of everyone it has trained over.

If we can train an AI over a small dataset or maybe even a large dataset of old art, or some mix in between, and then maybe fine tune it wtih a snall sampling of modern art, then I believe it would be unobjectionable, as this is largely how humans do it.

Why is it different?

The only difference that matters, is scale. And again, if I want to argue that something done 10000000000 times is legally problematic, I have to be prepared to explain why doing it 10 times is problematic as well, only less so.

I am completely aware that scale is a "thing" in legal systems. But as I said before: For scale to be important, the unscaled act in itself has to be problematic already.

Granted, things around power concentration have deep philosophical and social roots.

Could I have a dollar? What about a billion dollars?

The burden isn't on me to explain why something being scaled up by a billion is not the same.

ML can do either facsimile or imitation far better than a human mind can.

You seem to be conflating both things and suggesting that ML only does facsimile, which is where the potential legal problems are.

No, it is not. Memorization =! understanding.

I can teach a parrot to spew the times table, good luck getting it to understand how to apply it.

And a parrot is billions upon billions of times more capable than any current AI algos.

Diffusion would appear to me to work in much the same way. It doesn't understand what's good ("works", creates acceptable output) or why, but it knows it when it sees it, and has the tools to refine it.

AI, like humans, is capable of both imitation and facsimile. It is far superior at both feats.

Your fallacy is that you are noticing AI is superior at facsimile and erroneously assuming it is “not learning”. You are also ignoring the other amazing learning feats of imitation in front of you.

Parrots are lovely animals, but it’s unclear what you think you’ve accomplished by bringing them up. The fact that they are capable of more than just memorization does not differentiate them from advanced AI models, which are rapidly gaining all sorts of abilities.

Polly the parrot would have a hard time producing a picture of Elmo with a light saber in a Superman costume riding a dragon on the moon in the style of Rembrandt (in under 300ms, at least). I also know a parrot couldn’t write a 500 word story about the image.

I am pretty sure that GPT-3 is a lot more capable than a parrot in transpiling a function written in Python to Golang, or writing a summary to a tech-magazine article.

Same as StableDiffusion is ALOT more capable than me in drawing, painting and generally making up pretty pictures.

But there are no copies. For example, the LAION-2b training data is a total of 240 TB. The pruned SD model based on this dataset, is less than 5GB.

The data isn't copied into the models, it is used to teach the models, letting them learn patterns in the dataset.

Whether or not it's identical to human brains isn't the matter, they'd need to prove how a small 5GB model trained from a huge dataset infringes their rights specifically.

I have a 1.9 GB mp4 file on my harddrive. It contains 2 hours and 15 minutes of 1080p video data at 24 fps. Assuming it was generated from 4096x2160 16-bit color depth source material, the "training data" was 10.32 TB. I bet I could even get a similar size reduction as LAION-2b if I recompressed it to 720p.

Could I not also claim that I created an advanced AI model, which did not copy but learned patterns in the dataset? Modern video compression algorithms are getting quite complicated, after all.

I think no reasonable person would agree with this, but can you prove that the AI model is doing something substantially different?

Such patterns would enable the video file to decode into a multitude of pictures not originally in the training data. Obviously, a video file cannot do that...it's just compressed data.

Generative models however can generate things that are not in its training set.

And of course, there is a fundamental difference in the source data between compressed video and a generative model: video codecs work with a sorted sequence of images, where most images are slight variations of the ones before them. The training for generative AI doesn't have these properties, the input is not an ordered sequence, and even similar pictures are not sequential variations of one another.

Relying solely "uncompressed" size does not a really good metric make (this is analogous to the raw input size of the LAION dataset): one could make a reasonable argument that there are not billions (1) of image-pairs that are effectively identical up to a minute shift. I would posit the correct basis would be the Shannon entropy of the "best fit" ordering (minimize inter-frame diff), versus the lossy-compressed video, and a similar "best fit" ordering for the LAION dataset vs. the model.

My suspicion is that one will find that the relative number of "smooth transition" pairs in LAION viz the whole will be very different from the video.

-------

(1) - Napkin math: There are about 194400 frames, so ~37 billion (37791165600) frame-pairs. Assuming you have runs of about 1 second between hard cuts throughout, so an incidence rate of 1/24 for non-smooth transitions, gives us about ~36 billion "smooth transition" frame-pairs. I think it is safe to assume "on the order of" 1 billion, then. This ignores long "action" scenes with significant variance in images throughout, but also ignores longer-than-1-second slower scenes, hence the order-of-magnitude shrink in the assumption as buffer.

No, but what if you then produce "your own" rendition, or "remix", of that book, song or movie and offer it to the public? E.g. you memorize a collection of Taylor Swift's latest songs, and then start performing a medley of her hits in your local clubs, you may well find yourself in trouble.

But that never, ever, ever happens with current AI. it is not AGI. it is inspired by nothing, has no creativity, nada, ziltch.

Or it’s just too fast so let’s stop it?

Bear in mind, AI is not making artists or creative types obsolete - that would be fair game, just like computers made human calculators obsolete. No, this is about abusing other people's work.

If using someone else work for learning is infringement, then that's going to cause a lot of difficulty for all artists. Try making a rock song without listening to rock, or paint some modern art without viewing it etc.

The loophole is to use copyrighted works for free despite no learning taking place - no human being observing and developing their skills based on that work - rather, an algorithm transforming those works into some other useful interpretation of them.

[0]: https://en.wikipedia.org/wiki/American_Broadcasting_Cos.,_In.... [1]: https://www.vox.com/2018/11/7/18073200/aereo