What are the main risks and failure modes when deploying generative AI assistants for technical learning content at Pluralsight?