The Inventors of Deep Research

This Latent Space podcast episode interviews Aarush Selvan and Mukund Sridhar, the product and tech leads for Google's Gemini Deep Research. The discussion centers on Deep Research's functionality as a research assistant, its iterative research planning process (allowing user edits), and the technical challenges of building such an agent (e.g., balancing context window limitations with retrieval). Specific details include the use of a fine-tuned Gemini model and the surprising user preference for longer, more thorough research processes. The interview concludes with a discussion of future directions for Deep Research, including personalization and multimodal outputs. Listeners gain insights into the design and engineering of a cutting-edge AI research tool.