Unlock: Scaling Compute-Optimal Training

Chinchilla scaling: how to optimally allocate a fixed compute budget between model size and training data, why many models were undertrained, and the post-Chinchilla reality of data quality and inference cost.