Scaling Laws Glossary
Scaling laws are empirical relationships showing that, over certain ranges, AI model performance improves predictably as training compute, dataset size, and model size increase. They are one reason the AI labs invested so heavily in larger models and larger training runs: capability gains were not random; they followed curves.
The laws are not magic and not destiny. Data quality, architecture, post-training, tool use, inference-time compute, and product design all matter. Still, scaling laws gave the frontier labs a strategic map: spend more compute in the right way and the model will often get better before anyone fully understands why.
For the Dictionary, scaling laws sit behind the capability-overhang argument. If capability arrives before institutions, products, and users know what to do with it, the bottleneck shifts from invention to implementation.