Zhipu Founder Tang Jie on Scaling Law: Trillion-Parameter Large Models Were an Industry Detour, Next Key Lies in Post-Training

Zhipu founder Tang Jie stated that simply stacking trillion parameters was a detour the industry took, and the evolution of large models has shifted toward deep reasoning. Using GLM-5.3 as a controlled variable experiment, he demonstrated that with the base model and 753B parameters held constant, significant capability gains can be achieved through enhanced post-training. Parameter count is only one dimension of scaling; future breakthroughs require a comprehensive consideration of multiple variables, including inference cost, MoE activation mechanisms, and post-training.