Latent Space: Death of Params: Z.ai CEO Jie Tang on GLM 5.3 and the new Post-training Scaling Law
Z.ai CEO Jie Tang argues that parameter count is no longer a sufficient metric for evaluating model sizes, advocating for consideration of data, compute, and operational conditions. The article introduces GLM-5.3, which achieves significant performance gains through RL on long-horizon environments, simulating complex engineering and research workflows. This approach emphasizes models taking ownership of substantial work end-to-end, rather than relying on users for problem decomposition. The scaling process is supported by pipelines that synthesize environments and reward signals, with research agents collecting task patterns from real work. The discussion also covers the five knobs of scaling, including MoE sparsity, and notes that advanced skills require long causal chains rather than just total parameter count beyond a certain knowledge threshold. The article also touches on open-weight models, compression, agent harnesses, and RL systems work.
Original sources
Newsletter Latent Space
Issue:[AINews] Death of Params: Z.ai CEO Jie Tang on GLM 5.3 and the new Post-training Scaling Law ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.