Z.ai CEO Jie Tang: Post-Training Scaling Laws & Synthetic Environments

Latent Space · rss · 2026-08-20

Latent Space summarizes Z.ai CEO Jie Tang's insights on GLM 5.3 and post-training scaling laws.

Core Views:

Death of Parameter Count: Parameter count is only meaningful when combined with data volume, compute allocation, and deployment conditions.
Post-Training Scaling: In the inference era, memorization relies on parameters, while reasoning relies on post-training data and effective depth. GLM-5.3's gains come from RL in long-horizon environments.

End-to-End Synthetic Environments:

Research agents collect task patterns from real work and convert them into runnable environments with multi-step dependencies and hidden states.
A judge agent verifies solvability, and verifiers are synthesized without access to reference solutions to provide binary rewards.
This enables a "synthetic all the way down" recursive self-improvement pipeline.

Scaling Knobs: Tang identifies 5 dimensions of scaling, including MoE sparsity (XA-YB). Advanced skills (e.g., vulnerability finding) require maintaining long causal chains (20+ steps), not just total parameter count.

Industry Updates:

Model Release: Ornith-1.5 (MIT licensed) launches in 9B/35B/397B variants, focusing on end-to-end self-improvement.
Compression: UnslothAI releases 1-bit quantizations retaining 77% BF16 accuracy on 8GB RAM.
Agent Harness: DeepSeek Harness is positioned as an open agent runtime; TrueFoundry open-sources TrueForge, claiming 30% token reduction.

Original post →

Z.ai CEO Jie Tang: Post-Training Scaling Laws & Synthetic Environments

More from Companies & People