Introduction to Deep Research

Key Concepts

• Agentic Models: Models designed to perform tasks autonomously, thinking and planning over extended periods.
• Deep Research: An agentic model capable of multi-step research on the internet, discovering, synthesizing, and reasoning about content.
• Latency Constraints: Artificial limitations on the time a model can take to produce a response. Deep Research removes these constraints.
• O Series: OpenAI's line of reasoning models, designed to think for longer periods before providing an answer.
• Tool Use: The ability of a model to access and utilize external tools, such as web browsers or Python interpreters.
• AGI Roadmap: OpenAI's plan for developing Artificial General Intelligence, which includes models capable of autonomous knowledge discovery.
• Reinforcement Learning: A type of machine learning where an agent learns to make decisions by receiving rewards or penalties for its actions.
• Hallucination: The generation of incorrect or nonsensical information by a model.

Deep Research: An Agentic Offering from OpenAI

Introduction

OpenAI introduces Deep Research, a new capability designed to transform knowledge work by enabling multi-step research on the internet. This model aims to streamline processes, enhance worker productivity, and provide valuable assistance to consumers.

What is Deep Research?

Deep Research is a model that performs multi-step research on the internet, discovering, synthesizing, and reasoning about content. It adapts its plan as it uncovers more information. A key feature is the removal of latency constraints, allowing the model to take 5-30 minutes or even longer to produce an answer. This extended processing time is considered beneficial for autonomous task completion and aligns with OpenAI's AGI roadmap. The output is a comprehensive, fully cited research paper, similar to what an analyst or expert would produce.

Use Cases

• Knowledge Work: Streamlining research tasks, generating reports, and providing expert-level analysis.
• Specific Searches: Finding tailored items or information with specific constraints.
• Content Creation: Assisting in the creation of presentations and other content.
• Market Research: Analyzing market trends, adoption rates, and emerging opportunities.
• Academic Research: Assisting in various academic fields like physics, computer science, and biology.
• Product Research: Gathering information and reviews for potential purchases.

Deep Research in ChatGPT

Deep Research is integrated into ChatGPT via a dedicated button. Users can input a query, and the model will initiate the research process. The model may ask clarifying questions to ensure it understands the requirements accurately. A sidebar displays the model's reasoning process, showing the searches it conducts, the pages it opens, and the information it extracts.

Example: A product manager asks Deep Research to research iOS and Android adoption rates, the percentage of people who want to learn another language, and the change in mobile penetration over the past couple of years, comparing developed and developing countries. The model is asked to provide this information in a formatted report with tables and a recommendation on the best emerging opportunities for ChatGPT.

Under the Hood: Technology and Training

Deep Research is powered by a fine-tuned version of OpenAI's soon-to-be-released O3 reasoning model. It is trained using end-to-end reinforcement learning on hard browsing and reasoning tasks. The model learns to plan and execute multi-step trajectories, reacting to real-time information and backtracking when necessary. It can browse user-uploaded files, use a Python tool for calculations and image creation, and embed plots and images in its final response. Citations include specific sentences and passages from sources.

Performance and Evaluation

• Humanity's Last Exam: Deep Research achieved a new high of 26.6% accuracy on this benchmark, which tests capabilities across a range of expert subjects.
• Guia Benchmark: The model reached a new high on all three levels of difficulty on this benchmark, which measures agentic capabilities and requires web browsing, multimodal capability, code execution, and reasoning over files.
• Internal Expert-Level Evaluations: Experts rated the model's responses to tasks that would have taken them hours to complete. Pass rates were more correlated with estimated economic value than with the estimated number of hours to complete the task.
• Hallucination Evaluation: The model performs the best on this evaluation of any model OpenAI has released, but hallucination is still possible.

Examples of Deep Research in Action

• Investment Memo: The model generated a comprehensive investment memo analyzing the market for civilian supersonic air travel, using 12 different sources in 8 minutes.
• Biology Paper Search: The model found relevant papers on the same topic as a user-uploaded paper, which was validated by an expert.
• TV Show Identification: The model identified a TV show episode based on a vague description of its plot, demonstrating its ability to find "needle in a haystack" information.
• Ski Recommendation: The model recommended skis based on user preferences, including length, color palette, and intended use (all-mountain and powder).

Future Directions

OpenAI plans to expand Deep Research by connecting it to custom contexts and enterprise data storage. The goal is to develop agents that can think longer and more autonomously to solve very difficult tasks.

Conclusion

Deep Research represents a significant step towards more capable and autonomous agentic models. Its ability to perform multi-step research, synthesize information, and reason about content has the potential to transform knowledge work and assist users in a variety of tasks. While still in its early stages, Deep Research demonstrates the potential of AI to augment human intelligence and solve complex problems.