Scaling laws
Regular patterns showing that loss falls predictably as models, data and compute grow.
Also: neural scaling laws
In 2020 Jared Kaplan and colleagues at OpenAI showed that a language model's loss falls smoothly, following a power law, as parameters, training tokens and compute increase. That predictability let labs plan larger runs with confidence and helped justify GPT-3 and its successors. DeepMind's Chinchilla work in 2022 revised the best balance between model size and data. Since 2024, labs have also reported scaling laws for reinforcement learning and for compute spent at inference time. Whether these trends will continue is a matter of debate.