after you go from from millions of params to billions+ models start to get weird (depending on training) just look at any number of interpretability research papers. Anthropic has some good ones.
> just look at research papers
You didn't add anything other than vibes either.