Its relevant because those same issues occur with large models. Also, the "crappy small model", was a model trained for safety tasks, and outperformed the frontier lab safety models.