GLM 5.3 shows parameter count obsolete; scaling moves to data and compute

The shift Tang describes reflects a fundamental change in how the AI field understands model scaling. For years, parameter count served as a reliable proxy for capability—more parameters meant better performance. Chinchilla and subsequent scaling laws codified this relationship, but those assumptions broke down in what the article calls the "Inference Inflection world," where task-dependent token efficiency varies wildly (200–900 tokens per parameter). Tang's observation—that memorization favors more parameters while reasoning favors more post-training data and effective depth—explains why GLM 5.3 gained so much ground using the same architecture: the RL training environments were designed to teach long causal reasoning chains, not to expand memorization capacity.

This reframing has immediate practical consequences. Companies building or selecting AI systems can no longer rely on parameter count as a shorthand for comparing models. Instead, they need to evaluate training recipe quality, post-training strategy, and the conditions under which a model will operate. The broader trend visible across the article—from Ornith-1.5's synthetic environment generation to Microsoft's Agent Lightning and TrueForge's focus on the session and tool-orchestration layer—suggests the industry is redistributing effort away from model size and toward the infrastructure and training processes that unlock reasoning capability. For open-weight model developers, this creates both a challenge (building synthetic environments at scale is harder than adding parameters) and an opportunity (leaner architectures trained well can compete with larger ones).