This is true but I also find the opposite true: the model is highly dependent on the input/prompt, to the point that it can't think beyond it or about it.