Scaling Laws ☆

Scaling laws (Kaplan 2020; Chinchilla 2022) show model loss falling predictably as compute, parameters and data grow together — the smooth curves that convinced investors trillion-dollar infrastructure was rational. Chinchilla corrected the recipe (more data per parameter); test-time compute added a second axis; the open question — whether pre-training returns are flattening — is the trillion-dollar debate, with Ilya Sutskever's 'age of scaling is over' remarks its sharpest statement.