No Priors: Artificial Intelligence | Technology | Startups Inside Deep Research with Isa Fulford: Building the Future of AI Agents
667 snips
Apr 24, 2025 
Isa Fulford, a key figure in deep research at OpenAI, discusses the evolution of AI agents with real-world capabilities. She highlights the importance of using human expert data and how reinforcement learning enhances agent functionality. Isa delves into the challenges of latency, safety, and the balance between functionality and trust as AI systems become more autonomous. She also shares predictions on the future of AI agents, including their potential to complete tasks more efficiently and develop a sense of 'taste' in their operations.
AI Snips
Chapters
Transcript
Episode notes
When To Use Reinforcement Fine Tuning
- Use reinforcement fine tuning when your task is very different from existing data or critical for your workflow.
- Avoid it if model improvements on general use cases naturally keep progressing over time.
Human Expertise Guides Deep Research
- Expert human data was essential to model success because it guides the difficult information synthesis process.
- RL lets the model learn strategies from task goals without knowing the full expert research process.
Balancing Agent Trust and Safety
- Deep Research's long and comprehensive responses increase user trust but also risk more impactful hallucinations.
- Future agents need safe execution of tasks without unintended side effects to truly assist users.