GLM 5.3's biggest leap reportedly came from post-training alone

SemiAnalysis says GLM 5.3's biggest leap came from post-training alone and shares a quote saying the model was built entirely by post-training its predecessor. The quoted speaker describes a “night and day” difference in cyber capabilities compared with GLM 5.2 and argues that post-training is more GPU-efficient than pre-training from scratch. They say companies with proprietary data could use it to beat some closed models.