newsfilter
← Interviews
Sequoia Capital33 min

OpenAI’s Deep Research Team on Why Reinforcement Learning is the Future for AI Agents

Isa Fulford, Josh Tobin, Sonya Huang, Lauren Reeder

Interview, Fireside Chat, Product Demonstration
  • Reinforcement learning and fine-tuning are identified as critical components for building the most powerful agents, with the team projecting that 2025 will be the breakout year for the agent category.
  • The current AGI approach, defined as combining a reasoning model with human-like tools and outcome optimization, is viewed as ready to scale to complex tasks and a wide range of use cases.
  • Future agent products will expand beyond current offerings like Deep Research and Operator, including capabilities to browse, analyze, integrate information, and handle obscure or non-first-page internet data.
  • Models are expected to evolve to embed images and generate graphs in responses, while simultaneously improving citation accuracy and reducing hallucinations.
  • Data sources will expand to include private information alongside public data, enabling more comprehensive synthesis and analysis of existing knowledge rather than new scientific discovery.
  • Time-saving expectations include returning 5%, 10%, 20%, or 25% of user time on specific tasks, potentially reducing four-to-eight-hour workflows to five minutes.
  • A dual-path strategy is anticipated for workflows: hard rules encoded in human-written logic for predictable processes, and end-to-end trained models like Deep Research for scenarios requiring flexibility or handling many edge cases.
  • The team envisions shifts in education toward personalized learning tailored to user context, replacing traditional textbook methods with more engaging AI-driven instruction.
  • Key application areas identified include medical research (literature and clinical trials), business planning (market analysis and CPG startup), travel and shopping recommendations, and coding assistance.
  • Specific use cases range from complex planning (e.g., birthday parties) and detailed fact-finding (e.g., specific historical facts or TV show episodes) to market sizing and domain availability checks.