I don't think this aligns with the Chinchilla scaling law. There is indeed a point at which you can oversaturate a model with data, as well as such thing as not

I don't think this aligns with the Chinchilla scaling law. There is indeed a point at which you can oversaturate a model with data, as well as such thing as not giving it enough. Compute is the constraining factor, and it scales more-or-less linearly in both directions.