Troed Sångberg on Nostr: Did you know that the reason LLMs are trained on "all data they can get their hands ...

Did you know that the reason LLMs are trained on "all data they can get their hands on" is _not_ to copy that data or to be able to regurgitate it later?

It's actually the opposite. See, "overfitting" is when a model has too little training data to work with, and it memorizes question<->answer pairs. That causes the model to be unable to apply its knowledge on new scenarios it hasn't seen.

The solution is to train the model on much more data than it has a chance of memorizing. The model will then instead learn patterns, which it will then be able to apply to both existing as well as new problems, and offer up solutions.

Congratulations - now you know more than everyone who posts on social media that AI companies "steal everyone's data".