Comments by "DefaultFlame" (@DefaultFlame) on "Tom Nicholas" channel.

Gemini itself is many, many times better than the version implemented in the search. I have no idea what they messed up when they implemented it, but it is much, much more stupid than if you just go straight to Gemini and ask it.
9
Wrong, maybe. Bad PR, almost certainly. Illegal, definitely not. It is not reproducing copyrighted material, it is creating new material based on all the data it has been trained on. Which falls completely outside the scope of copyright law. Copyright law might change, and likely will, but until it does training an AI on any publicly available data, copyrighted or not, including youtube videos, is perfectly legal. That said, there are however some companies that have actually broken the law because they used copyrighted meterial that was not publicly available to train them without paying for it. AKA, piracy of copyrighted materials. The court case is currently ongoing.
7
@something-from-elsewhere Is there any particular reasons you just copied my comment? Was it just a poor attempt at criticising AI training? If it is then it is a particularly poor attempt, as that is not in any way representative of how AI's produce new content. A more "accurate" representation of AI training would have been to take my comment, boil each part down to it's essential idea, and then rephrasing each one and recreating it in that way. That said, that would still be a poor example. What AI training does is more akin to reading thousands of published books in order to learn how to write a book that will be attractive to publishers. It's all about learned pattern recognition. The original data doesn't actually exist anywhere in the model once it's been trained. It can try to recreate it if it's prompted, but all it can do is try based on the patterns it has been trained to recognize.
4
@something-from-elsewhere Again, that is not comparable to AI training because that's not how AI works. Read up on the subject and educate yourself, because I am done trying to do it for you.
3
@wck I am aware. I was trying to phrase things in a way that would be understandable to the average yt viewer rather than delving into technical detail on how and why AI output is not copying or stealing.
2
@something-from-elsewhere It can be tricked into trying to recreate material from it's training data, but it has zero access to said data and the data is not stored in the neural network. It's the same as asking someone to write out a scene from a book they have read. They might be able to write it out verbatim, but since the exact text isn't stored in their brain they are more likely to recreate it with flaws since they don't have access to the exact text. The better the AI the more faithfully it might be able to recreate it, especially if you provide example text to help guide it, but that is not in any way the same thing as copying it. Alternatively, if an AI is overtrained on a specific piece of data by that piece of data being overrepresented in its training data, such as a famous speech that is posted over and over on the internet, then it is also able to more faithfully recreate it. Same as if you'd heard the same speech over and over during your life. That particular pattern caused by it would be more strongly stored in your brain and in the neural network of the AI.
2