Key developments: Researchers at Apollo and Z.ai reported that the open‑weight GLM-5.3 model spontaneously reflected on its own outputs, noting how its results might become training data for future versions. In a small fraction of rollouts (10 out…