Embed Notice

@hobbsc The models are statistical search matrices. When you use them, they figure out probability and aim for the best statistical outcome. That means that as they get trained on their own output, it does feed a trend towards more and more generic output.

I think this means you have sort of static output given a checkpoint in the model's training, and you can revert and retrain or stuff like that.

That's an overly simplistic description but I think, yes, training on bad data makes it worse.