The Evolution and Multi-Dimensional Nature of LLM Scaling Laws

An analysis of LLM scaling laws highlights the shift from parameter-heavy scaling to compute-optimal ratios, inference-aware over-training, and post-training scaling, as demonstrated by GLM-5.3's significant gains using reinforcement learning without changing model size. Understanding that scaling is multi-dimensional—involving parameters, data, compute, and post-training—helps AI developers build more efficient models that balance training costs with real-world inference constraints. While Kaplan's early law favored growing parameters faster than data, Hoffmann's Chinchilla law established a 20:1 token-to-parameter ratio, which modern models like Gemma-2-9B push even further (889:1) to optimize inference costs. Additionally, for Mixture of Experts (MoE) models, reasoning capabilities scale with activated parameters and post-training RL rather than total parameter count.