The Inventors of Deep Research

Episode Summary

Executive Summary: The episode explores Google Gemini Deep Research with its PM and tech lead, focusing on how the product works, why it needed a specialized post-trained model, and how it differs from generic search or RAG. The guests explain the planning, browsing, verification, and report-writing loop, discuss UX choices like editable plans and side-by-side chat, and outline evaluation, latency, and future directions such as personalization, multimodal outputs, and deeper integration with user docs.

Main Topics: What Deep Research is and why it matters: Deep Research is positioned as an agentic research assistant that helps users go from zero to a solid understanding of a topic in minutes, especially when the question is broad, multi-faceted, or requires synthesis across many sources. Planning, browsing, and report generation workflow: The team explains the product’s internal loop: generate a research plan, let users edit it, browse the web iteratively and in parallel, then synthesize a grounded report with citations and follow-up support. Model specialization and post-training: They emphasize that Deep Research is not just standard Gemini with search turned on; it uses a post-trained version of Gemini 1.5 Pro tuned for planning, reliability, and research behavior. UX design for transparency and steering: The conversation highlights why the product shows the plan, exposes sources in real time, and uses a side-by-side artifact/chat layout so users can steer the research and continue iterating. Evaluation, ontology of use cases, and latency trade-offs: The guests describe how they evaluate a high-entropy product through a use-case ontology, human evals, and behavioral metrics, while balancing the value of longer research against user impatience. Context, RAG, and long-horizon memory: They discuss when to keep information in context versus move it into retrieval, how long-context models change the design space, and why recent turns should stay in memory for follow-up questions. Future directions: personalization, multimodality, and domain-specific research: The team sees opportunities in personalized reports, multimodal outputs, deeper integration with user documents and subscriptions, and more interactive agent workflows for complex projects.

Key Arguments: Deep Research is most valuable when a query has multiple facets and would otherwise require opening dozens of tabs and synthesizing information manually. The product deliberately shows a research plan before execution so users can correct scope and steer the agent, rather than forcing them to guess the right follow-up questions. The system uses iterative planning and browsing, with the model deciding when to search, when to double-click into sources, and when to stop based on accumulated evidence. A specialized post-trained model was needed because generic Gemini behavior was not sufficient for consistent research planning and grounded synthesis. Long context is preferred for recent turns and active threads, while older material can be moved to retrieval; the choice depends on how likely future comparisons are. Latency is a feature, not just a bug: users often value visible effort and thoroughness, even though the team initially expected faster responses to win. Benchmarks are useful for motivation and industry comparison, but they do not fully capture the real product experience of deep research. Future improvements should focus on personalization, multimodal reports, and access to private or subscription-based corpora, not just open-web search.

Data Points: Research time: about 5 minutes - Deep Research browses the web for roughly five minutes before producing a report. Alternative research time: 5 or 6 minutes vs. a week or two - Tyler Cowen’s quoted comparison of Deep Research to a PhD-level assistant. Hard stop expectation: never more than 10 minutes - The team said they initially insisted the product should not exceed 10 minutes. Context window: 1 million to 2 million tokens - They referenced very long context capacity and discussed what happens when it is exceeded. Initial website count: 36 - During the demo, the model initially surfaced 36 websites to explore. Launch date: February 2 - OpenAI’s Deep Research launch date referenced in the introduction. AI Engineer Summit date: February 21 - The guests were announced as keynote speakers for the summit in New York City. Use-case spectrum: broad-to-deep ontology with several midpoints - They described an ontology ranging from shallow option exploration to deep topic research, plus compound project-like tasks.

Pivotal Quotes: "Deep Research is the best public-facing AI product Google has ever released." — Jason Calacanis: Quoted in the introduction as an early reaction to OpenAI’s Deep Research. "I have had Deep Research write a number of 10-page papers for me, each of them outstanding." — Tyler Cowen: Quoted in the introduction to illustrate strong user enthusiasm for the category. "I think there’s definitely some things we can do on the UX side to basically invite the user to be like, Hey, this is the starting point. Now let’s keep going together." — Arush Selvan / Mukund Sridhar: On making Deep Research feel like an ongoing collaborative workflow rather than a one-shot report generator.

Implications: Deep research agents are becoming a real product category, but winning will depend on transparency, steering, grounding, and workflow integration—not just raw model intelligence. The next leap is personalized, multimodal, domain-aware research that can continue across tasks and private data.