Session brief
How Deep Research Works - Mukund Sridhar & Aarush Selvan, Google DeepMind
Mukund Sridhar , Aarush Selvan
Overview
This talk details the development of Deep Research, a personal research agent integrated into Gemini Advanced. The core thesis is that by removing compute and latency constraints at inference time, AI models can perform extensive web browsing to generate comprehensive answers, addressing the limitations of traditional chatbots that often provide blueprints rather than direct solutions. The presentation highlights the product and technical challenges encountered in building such an asynchronous, long-form research tool within a synchronous chatbot interface.
Who should watch
- AI Engineers
- Product Managers
- Builders working on agentic systems
- Those interested in long-context AI applications
- Individuals facing challenges in synthesizing information from the web
Key takeaways
- Deep Research was motivated by the need to provide comprehensive answers to complex research queries, moving beyond the "blueprint" responses of standard chatbots.
- Building Deep Research involved overcoming product challenges such as integrating asynchronous tasks into a synchronous chatbot and managing user expectations for long-wait times.
- Technical challenges included robustly handling long-running tasks, iterative planning, effective compute/time management, interacting with a noisy web environment, and managing growing context sizes.
- The user experience incorporates a research plan that users can edit, real-time browsing transparency, and a long-form report inspired by Anthropic's artifacts, allowing users to ask questions about the material.
- Transparency is maintained by showing all sources read and used, with options to export citations to formats like Google Docs.
- Future directions for research agents include developing deeper expertise beyond aggregation, personalizing output based on user needs (e.g., strategic positioning for VCs vs. financial modeling for bankers), and integrating multimodal capabilities like coding and data science.
- The system must be robust to intermediate failures, employing state management and error recovery for long-running research tasks.
- Users can engage with research results through interactive elements, follow-up questions, and the ability to steer the research direction.
Notable quotes
*We wanted to help people get smart fast.*
*What if you remove the constraints of compute and latency at inference time, let Gemini take as long as it wants, browse the web as much as it needs, and see if we can trade that off for a much comprehensive answer for the user.*
*We saw that research and learning queries are some of the top use cases in Gemini.*
Unofficial community note. Prefer the recording for nuance.