Return guests Kanjun and Jonathan discuss Imbue and Databricks' 70B LLM, outperforming GPT-4o. Topics include infrastructure needs, cost-aware hyperparameter optimizer CARBs, and challenges in training large models. They delve into MFU monitoring, bug tracing, and optimizing infrastructure for deep learning. The podcast explores evaluation metrics and AI model practical applications, emphasizing the importance of creating useful tools for customers.