Scaling requires linear co-scaling of three ingredients

Improving model performance requires linearly scaling up bigger networks, longer training times, and more data together — like a chemical reaction with three reagents; if you scale up one without the others, you run out of the other reagents and the reaction stops.