Maybe we can come up with smaller models that perform almost as well as the bigger ones. Could that just be pca of some kind?

Gpt nano vs gpt 5 for example.