3 ms·
An overlooked value of these language models is that they circumvent copyright. Google also did this when it appeared on the scene. It essentially sucked value
by leobg 4y ago
An overlooked value of these language models is that they circumvent copyright.
Google also did this when it appeared on the scene. It essentially sucked value from websites that, prior to search engines, would have been considered copyrighted works. But Google made itself the go-to place on the internet. In people‘s minds, all the content was „on Google“.
Now GPT can do the same thing. It can consume all kinds of text, articles, books, video transcripts. And can then present itself as the „go to“ place for finding answers.
Of course, copyright as such has always been an artificial construct. It has never protected ideas themselves - only the specific form of their expression. By being able to suck ideas (which are free) out of their specific form of expression (which is protected), GPT can essentially resell the value that previously was there to be monetized by those who held the copyright thereto.
In the human world, we had the same thing in the form of experts. People who read lots of books over the course of their life, and thus were able to answer questions, solve problems and write books of their own. But, of course, GPT can read more books than any human ever could. And it can answer more questions than any human ever can.
- leobg 4y agoI predict that content owners will push for legislation that makes using a copyrighted as training data for a neural network explicitly illegal without the owner‘s consent. They might argue that using copyrighted works to train neural networks was illegal even by the standards of today‘s law. But they will try to push to make it explicit, to avoid the obvious debate that copyright has never been designed as a tool to keep ideas hostage. And so I predict that the big winners of tomorrow are being created today, while there is still a certain Wild West and regulation is still foggy. Today, you can still train your models on Reddit, Common Crawl, and, who knows, maybe those Z-Library torrents that are out there? As long as your model will not quote its sources verbatim, it‘s very unlikely you‘ll face any consequences. But once the training data market becomes more regulated, it will become much more difficult to build a huge model without paying millions in license fees. In fact, I wouldn‘t be surprised if it OpenAI itself would not be pushing for more regulation at some point. Just like Google, once their own position seemed secure, began helping webmasters to „protect themselves“ against other people‘s automated crawlers (e.g. ReCaptcha).