AI & Enterprise
Rethinking AI performance scaling beyond parameter count
Jie Tang, a Tsinghua University professor and founder of AI startup Z.ai, said AI model performance is not determined by parameter count alone. He said data volume, how computing resources are used, and operating conditions also matter. Tang pointed to shifts from Jared Kaplan’s 2020 scaling results to DeepMind’s 2022 Chinchilla law and argued that inference-driven lifetime costs can favor smaller models trained longer. He also described how MoE architectures change scaling trade-offs.
