Was GLM-5.2 trained on Opus 4.5 outputs?
Zhiyuan Hu, a researcher at GLM University, has published a study claiming that the GLM-5.2 language model was likely trained on outputs generated by Anthropic's Claude Opus 4.5, based on statistical signatures and stylometric analysis of response patterns.
Background
- GLM-5.2 is a large language model (LLM) developed by Zhipu AI, a leading Chinese AI company. Opus 4.5 is a version of Anthropic's Claude model, one of the most capable Western LLMs.
- The article investigates whether GLM-5.2 was trained on synthetic data generated by Opus 4.5 — a practice known as "distillation" or "model imitation" where a weaker model is fine-tuned on outputs from a stronger one, often to shortcut the expensive training process.
- This matters because: (1) it raises questions about the true originality and capability of GLM-5.2; (2) distillation is common but often denied; and (3) it has competitive and ethical implications for the AI industry, especially given US-China tech tensions and export controls on advanced AI models.
- The technical clue: if GLM-5.2 closely mirrors Opus 4.5's distinctive error patterns, stylistic quirks, or refusal behaviors that are unlikely to arise independently, that suggests training on Opus outputs. The article likely analyzes such fingerprints.