Zhipu's Tang Jie on Scaling Law: GLM-5.3 Gains Come from Post-Training, Not Parameters

On August 19, Tang Jie, Zhipu co-founder, chief scientist and Tsinghua professor, published "Thoughts About Scaling Law" on X, presenting GLM-5.3 as a controlled experiment: it shares the same base, architecture, total and active parameters as GLM-5.2, with one month of scaled long-horizon environments and RL post-training yielding non-marginal gains—evidence that scaling knobs exist beyond parameter count.

Confirmed

• Tang argued that parameter count cannot be discussed in isolation from training data volume, compute allocation, and deployment conditions; the post also cited learning rate and other variables.
• He reviewed the shifts in scaling strategy: Kaplan (2020) concluded parameters should grow faster than data (2.7:1), pushing the industry toward parameter over-expansion; Hoffmann/Chinchilla (2022) revised the compute-optimal ratio to 20:1 (data:parameters) with both growing at similar rates; once inference cost dominates, the trend shifts toward smaller models trained longer (e.g., Llama-2), with inference and MoE reshaping the optimal ratio.
• He holds that total parameters matter until a threshold sufficient to contain the "world," after which gains come from scaling effective depth per forward pass and, most importantly, post-training.
• GLM-5.3 was obtained with identical parameter configuration, scaling only long-horizon environments and RL post-training; officials describe the gains as non-marginal.
• OpenAI researcher Liam Fedus reshared the post and added historical detail: in 2020, while exploring Switch Transformer sparsity, the team routed each token to 1 of 2048 experts, echoing the intertwined evolution of MoE and scaling.

Why it matters

• This is an explicit statement from a leading Chinese lab executive against "parameter worship," hinting future releases may not prioritize larger parameter counts.
• GLM-5.2 vs GLM-5.3 forms a rare controlled comparison isolating post-training, offering a reference for evaluating the scaling value of RL post-training.
• natolambert endorsed the view that parameters matter only alongside data, compute allocation and operating conditions; samsja19 extended the discussion to independent scaling targets for active vs. total parameters, depth-compute trade-offs, and test-time compute versus parameter scaling.

2026-08-19 ~ 2026-08-20 · 10 related posts

Full story(2 episodes)→

• Episode 1: Zhipu's Tang Jie on Scaling Law: GLM-5.3 Gains Come from Post-Training, Not Parameters(2026-08-19, 10 posts)
• Episode 2: Debate Rages Over Scaling Laws as Pretraining Potential Called Far From Saturated(2026-08-21, 3 posts)

• [source] Zhipu's Tang Jie on Scaling Laws: Inference and MoE Shift Optima — jietang · 2026-08-19
• [source] Z.ai's GLM-5.3 Essay: Scaling Law Has More Dials Than Parameter Count — pmttyji · 2026-08-19
• Zhipu's Tang Jie: Parameter Count No Longer the Sole Scaling Metric — 量子位 · 2026-08-19
• Discussion on Scaling Laws: Parameters, Depth, and Test-Time Compute — samsja19 · 2026-08-20
• Tang Jie on Scaling Law: Optimal split between data and compute — AccBalanced · 2026-08-20
• Scaling Law is About More Than Just Parameter Count — natolambert · 2026-08-20
• Liam Fedus on Scaling Laws: FLOPs are Intelligence, Parameters are Knowledge — hsu_byron · 2026-08-20
• Tang Jie: Trillion Params Was a Detour, Scaling Laws Shift — 青稞AI · 2026-08-20

2 near-duplicate retellings: 智东西 · AccBalanced