The finding that for a fixed budget there is a best balance between how big a model is and how much data it sees, and that earlier models were far too large for the amount of data they were trained on. It changed the industry’s direction from ever-bigger models towards smaller models trained much longer.
Technical view
The Chinchilla result established that parameters and training tokens should scale roughly in proportion, contradicting earlier scaling prescriptions. Note that compute-optimal minimises training cost for a given loss, and is not inference-optimal: since inference cost recurs on every request, production models are deliberately trained well past the compute-optimal point on far more tokens, trading training expense for a permanently cheaper model to serve.
Related
Part of AI Terminology Index, section: Training and Post-training.