This episode explores the development and capabilities of OpenAI's Deep Research, an agentic feature within ChatGPT. Isa Fulford, the lead of OpenAI's Deep Research team, details how this feature conducts multi-step online research to solve complex tasks, producing cited reports in minutes that would take humans hours. Against this backdrop, Fulford explains the development process, which involved initially creating a demo to generate excitement, followed by the creation of reinforcement learning tasks and tools (like browser access and code execution) to train the model. More significantly, the discussion highlights Deep Research's applications across professional fields (academics, VCs, consultants) and personal uses (shopping, travel recommendations), illustrated with examples like generating reports on venture capital investments in AI and finding nearby night markets. For instance, a live demo showcased Deep Research's ability to answer complex queries, synthesize information from various sources, and even generate graphs. In conclusion, while acknowledging limitations like occasional hallucinations, Fulford emphasizes ongoing efforts to improve reliability, integrate Deep Research into core models, and expand its capabilities to include private context and action-taking, signifying a significant advancement in AI-powered research tools.
