Scaling laws are empirical relationships showing how model performance improves predictably as you increase parameters, data, or compute — used to plan training runs before committing resources. These laws, established by research from OpenAI and DeepMind, enable teams to predict the performance of much larger models from smaller experimental runs.