AI training data has a price tag that only Big Tech can afford
AI training data has a price tag that only Big Tech can afford
Generative AI improvements are increasingly being made through data curation and collection — not architectural — improvements. Big Tech has an advantage.
All this data was "crowdsourced" -- i.e., stolen -- from the public in the first place. As far as I'm concerned, they owe us and have no room to complain if we steal it right back.
I recently listened to this German language podcast episode about the social cost and how life is for a few clickworkers in Africa: Das Wissen | SWR: Clickworker – Ausgebeutet für künstliche Intelligenz
Based on the post title alone, I call bull because I could buy enough storage and pirate enough books in order to create an AI, using copyrighted material as the training data. Yes it would be an absolutely horrible AI since I don't have a clue what I'd be doing, but it's possible.
Then go ahead and buy 2000 Nvidia cards.
The training data is important, but currently the bottleneck is computing power. Buying so many chips and having them run full blast 24/7 costs a lot of money.
You can get your hands on books3 or any other dataset that was exposed to the public at some point, but large companies have private human-filtered high-quality datasets that perform better. You're unlikely to have the resources to do the same.
It's not clear if this is piracy. In the US, it's obviously an ongoing fight. Basically, what you describe is "books3", put together with scripts by Aaron Swartz.
It's legal in Japan, if the purpose is only AI training and not enjoyment. I'm not sure if there are issues regarding DRM or such.
In the EU, the dataset and resulting model would be illegal. Any business offering the model would be in hot water, but I think internal use would be fine.