newsfilter.io
Conference Presentation, Product Demonstration

How OpenAI Built its Groundbreaking Deep Research Product ft. Isa Fulford

  • Deep research agents are expected to spend 5 to 30 minutes analyzing multiple online sources to generate fully cited, comprehensive reports comparable to those of a research analyst.
  • Models will be trained directly on daily user tasks, utilizing online browsing as an initial sandbox due to the constrained safety profile of read-only operations.
  • Deep research capabilities are planned for integration into main reasoning models and will likely upstream into larger models such as O3.
  • Future development will prioritize improving reliability to address current instances of model hallucination.
  • The roadmap includes integrating private context, specifically internal company knowledge and paywalled sources, into the deep research workflow.
  • Strategic focus is shifting from synthesizing existing information to executing actions, marking the next major development phase.