SC
How OpenAI Built its Groundbreaking Deep Research Product ft. Isa Fulford
@ Sequoia Capital
05/08/25
Takeaways
IF
Isa Fulford
05/08/25
@ Sequoia Capital
Deep Research is an agentic capability in ChatGPT that conducts multi-step online research to solve complex tasks, delivering a fully cited comprehensive report in minutes, which would take a human many hours.
IF
Isa Fulford
05/08/25
@ Sequoia Capital
We trained a version of GPT-3 specifically for web browsing and data analysis to enhance the capabilities of Deep Research.
IF
Isa Fulford
05/08/25
@ Sequoia Capital
We started building Deep Research over a year ago, driven by progress in reinforcement learning and reasoning models, initially focusing on math, science, and coding tasks.
IF
Isa Fulford
05/08/25
@ Sequoia Capital
We aimed to create a useful model by training it directly on tasks that users perform in their daily lives, with online browsing being a key focus due to its broad applicability.
IF
Isa Fulford
05/08/25
@ Sequoia Capital
To develop Deep Research, we created reinforcement learning tasks to teach the model browsing and data analysis capabilities, and provided it access to tools for training.
IF
Isa Fulford
05/08/25
@ Sequoia Capital
Most users of Deep Research utilize it for professional purposes, including academics and consultants, but it also serves personal use cases like shopping and travel recommendations.
IF
Isa Fulford
05/08/25
@ Sequoia Capital
Deep Research begins by asking clarifying questions to ensure it understands the user's request and wants users to provide as much detail as possible upfront to ensure the output meets their expectations.
IF
Isa Fulford
05/08/25
@ Sequoia Capital
Deep research can kick off a research task and provide a summary of its reasoning and actions taken during the process.
IF
Isa Fulford
05/08/25
@ Sequoia Capital
The model will show an interleaved chain of thought, reasoning through the sources it encounters and the tool calls made, such as searching or analyzing with Python.
IF
Isa Fulford
05/08/25
@ Sequoia Capital
Deep research is capable of handling complex queries that traditional search methods struggle with, such as finding specific markets based on multiple constraints.
IF
Isa Fulford
05/08/25
@ Sequoia Capital
Deep research can effectively search the web, identify items that fit specific constraints, and synthesize the information into a comprehensive report.
IF
Isa Fulford
05/08/25
@ Sequoia Capital
The final response from deep research includes a detailed report with suggestions and citations from relevant sources.
IF
Isa Fulford
05/08/25
@ Sequoia Capital
The model summarizes its thought process, detailing its plans and actions taken during the search.
IF
Isa Fulford
05/08/25
@ Sequoia Capital
Deep research is designed to be the most comprehensive tool for search queries, while the 03 model is better suited for medium-level searches.
IF
Isa Fulford
05/08/25
@ Sequoia Capital
The 03 model is effective at searching because it is trained with the same tools and browsing datasets used for deep research.
IF
Isa Fulford
05/08/25
@ Sequoia Capital
Deep research can provide insights across a broad range of topics, including specific gene therapies that have gained regulatory approval in the US.
IF
Isa Fulford
05/08/25
@ Sequoia Capital
We're excited to continue developing deep research, improving its reliability, and integrating it into our main reasoning models.
IF
Isa Fulford
05/08/25
@ Sequoia Capital
Future developments for deep research include incorporating private context and enabling it to take actions based on synthesized information.
IF
Isa Fulford
05/08/25
@ Sequoia Capital
Thank you for having me, and I'll pass over to Zach.