AI summary by TwinMind
How OpenAI Built its Groundbreaking Deep Research Product ft. Isa Fulford
Sequoia Capital
8m 55s
Introduction & Speaker Background
- Issa Fulford leads Deep Research team at OpenAI
- Traveled from Korea specifically for event, landed 5:45 AM same morning
- Deep Research: agentic capability in ChatGPT for multi-step online research
- Solves complex tasks in 5-30 minutes vs many human hours
- Powered by fine-tuned O3 version specialized for web browsing & data analysis
- Produces fully cited comprehensive reports at research analyst level
Deep Research Development History
- Started 1+ year ago during internal RL & reasoning model progress
- Initial training focused on math, science, coding tasks
- Saw generalization to other domains but wanted direct training on daily user tasks
- Chose online browsing as first target:
- Broad domain coverage across jobs & daily life
- Good readonly agent sandbox with constrained safety considerations
- Team: Issa Fulford, Yash Patil, Thomas Simpson
- Initial approach: hacked demo using model prompting (no training)
- Generated excitement, then built proper training pipeline
Technical Implementation
- Created RL tasks for browsing & data analysis capabilities
- Provided tools during training:
- Browser access (search, click, scroll)
- Code execution for data analysis & graph plotting
- Fine-tuned O3 specifically for web browsing performance
- O3 base model benefits from same browsing datasets/tools developed for Deep Research
Product Design & User Experience
- Starts with clarifying questions to ensure precise output
- Encourages user specificity upfront given multi-minute processing time
- Shows interleaved chain of thought + tool calls during execution
- Provides detailed citations linking to specific source lines
- Takes 5-30 minutes for comprehensive research tasks
Use Cases & Examples
Professional Applications
- Academics, VCs, consultants
- AI venture capital investment analysis (live demo)
- Recent trends, funding patterns, visual graphs
- Generated comprehensive report with unusual March spike (likely OpenAI investment)
- Gene therapy regulatory approval research
- Specific hemophilia treatments with US regulatory approval
- Correct citations & explanations
Personal Applications
- Shopping & travel recommendations (speaker's primary use case)
- Korea night market search example:
- Maximum 15 minutes from location
- Reddit + Korean sources analysis
- Best rated stores within each market
- Successfully used recommendation in real life
Current Limitations & Future Development
Current Issues
- Occasional hallucinations
- Reliability improvements needed
Planned Enhancements
- Integration into main reasoning models
- Private context integration (internal company knowledge, paid sources)
- Action-taking capabilities beyond information synthesis
- Upstream improvements to larger reasoning models
