Putting the two papers together

These papers are not in conflict. The first one discovered that scaling works in a predictable way. The second one refined the recipe for how to scale, by balancing model size against data instead of just inflating the model.

Question	Scaling Laws (2020)	Chinchilla (2022)
Does scaling help predictably?	Yes, it follows a smooth power law	Agrees
Where to spend extra compute?	Mostly on a bigger model	Split evenly between model size and data
Practical rule	Bigger is better	About 20 tokens of data per parameter
Lasting lesson	Forecast performance before training	Do not starve your model of data