Scaling laws describe how a language model improves as the model size, the amount of training data and the compute budget grow.

Single Important Point

  • for a fixed compute budget, the model size and the number of training tokens must be balanced
  • a very large model trained on too few tokens wastes compute
  • a small model trained on a huge number of tokens also wastes compute

What it rules out

  • always maximising model size is wrong
  • always maximising data is wrong
  • the learning rate is not the lever here

The models these laws were measured on are in Generative Pre-trained Transformer (GPT).