Introduction to Deep Research

In the video, Mark from OpenAI introduces Deep Research, a new capability that enhances AI agents’ ability to conduct multi-step research on the web, allowing them to synthesize and reason about information over extended periods. Demonstrations highlight its effectiveness in generating comprehensive reports and tailored recommendations, showcasing its potential to transform knowledge work and contribute to the development of artificial general intelligence (AGI).

In the video, Mark, the head of research at OpenAI, introduces a new capability called Deep Research, which aims to enhance the functionality of AI agents in knowledge work. Joined by team members Issa, Josh, and Neil, they discuss the significance of agents in transforming both enterprise processes and consumer experiences. The introduction of Deep Research follows the launch of the O1 model, which was designed for reasoning but lacked the ability to access external tools like the internet. Deep Research addresses this limitation by enabling multi-step research on the web, allowing the model to discover, synthesize, and reason about information over extended periods.

Deep Research is characterized by its ability to take longer to return results, sometimes up to 30 minutes, which is seen as a positive feature. This extended processing time allows the model to perform autonomous tasks more effectively, aligning with OpenAI’s roadmap towards artificial general intelligence (AGI). The ultimate goal is to create a model capable of uncovering and synthesizing new knowledge independently. The output from Deep Research is akin to a comprehensive, fully cited research paper, similar to what an expert or analyst would produce.

Neil demonstrates how Deep Research works within ChatGPT by initiating a query related to market research for a language translation app. The model first asks clarifying questions to ensure it understands the requirements before beginning its research process. As it conducts searches, it opens various web pages, synthesizes the information, and adapts its approach based on the data it uncovers. This interactive process showcases the model’s ability to think critically and adjust its research trajectory in real-time.

Josh highlights another use case for Deep Research, focusing on personal purchases, such as buying skis while in Japan. He emphasizes the model’s capability to format its output as a report, complete with tables and recommendations. The model’s ability to ask clarifying questions ensures that it gathers the necessary information to provide tailored recommendations. This feature is particularly useful for consumers looking to make informed decisions without spending excessive time on research.

Issa explains the underlying technology of Deep Research, which is powered by a fine-tuned version of the upcoming O3 reasoning model. The model has been trained using reinforcement learning to execute complex tasks that typically require significant human effort. Internal evaluations show that Deep Research performs exceptionally well on expert-level tasks, demonstrating its potential to revolutionize knowledge work. Mark concludes by reiterating the importance of Deep Research in OpenAI’s vision for AGI, emphasizing its ability to handle complex tasks autonomously and efficiently.