Zhipu AI CEO Jie Tang on GLM 5.3 and Post-Training Scaling Laws
Zhipu AI CEO Jie Tang discussed the release of GLM 5.3, a flagship Mixture of Experts (MoE) model with a 1-million-token context window, highlighting a shift in AI development. He emphasized that the industry is moving away from traditional pre-training parameter scaling toward post-training scaling laws. This shift suggests that raw model size (parameter count) is no longer the sole metric for AI progress, as post-training techniques like reinforcement learning and inference-time reasoning yield better efficiency and capabilities. It signals a new phase of AI infrastructure optimization where compute is allocated during inference and reasoning rather than just initial training. GLM 5.3 features a 1-million-token context window designed for advanced text generation and deep reasoning. The model leverages Mixture of Experts (MoE) architecture, building upon the foundations of previous GLM models to optimize performance.
## BACKGROUND
Traditionally, AI progress relied on pre-training scaling laws, which state that LLM performance improves predictably by increasing dataset size, compute power, and parameter count. Recently, the industry has embraced post-training scaling (using reinforcement learning) and test-time scaling (allowing models to 'think' longer during inference) to achieve frontier-level reasoning without exponentially growing model sizes.