Zhipu's Tang Jie on Scaling Laws: Inference and MoE Shift Optima

jietang · x · 2026-08-19

Zhipu's Tang Jie provides a deep analysis of Scaling Law evolution. He notes that early Kaplan laws led to over-expansion of parameters, while Chinchilla corrected the compute-optimal ratio to 20:1. However, as inference costs dominate, the trend has shifted to smaller models trained longer (e.g., Llama-2).

MoE architecture further changed the target: total parameters govern knowledge retention, while activated parameters govern reasoning depth. Roberts (2025) found the optimal ratio is task-dependent, and pushing total parameters too high can actually degrade reasoning.

GLM-5.3 serves as a controlled experiment: keeping the base and parameters constant, it achieved non-marginal gains through one month of long-horizon RL and post-training, proving that beyond a parameter threshold, effective depth and post-training are key.

Original post →

More from Models

Models channel →