This episode explores the development and capabilities of OpenAI's Deep Research, a new agentic product released in February. Isa Fulford, a key developer, details its origins, stemming from internal advancements in reinforcement learning algorithms initially applied to math and coding problems. The team then shifted focus to more user-centric tasks, such as online research and information synthesis, aiming to create a tool useful for knowledge workers and aligning with OpenAI's broader AGI goals. For instance, the model was trained to find all papers co-authored by specific researchers and even locate a coworker's middle name. More significantly, the discussion delves into the challenges of data creation, tool development (a text-based browser with access to PDFs and Python tools), and the complexities of reinforcement fine-tuning. The conversation also touches upon safety concerns, the potential for hallucinations, and the future of agents capable of both research and action-taking. Ultimately, this highlights the rapid progress in AI and the emerging potential for agents to become increasingly sophisticated and integrated into various workflows.
