OpenAI announces that ChatGPT will have a 'Deep research' feature to collect online information
OpenAI has announced that it will introduce a ' Deep research ' feature to ChatGPT that allows AI models to collect online information. Deep research is an AI agent based on OpenAI's inference model 'o3', which can search for information on the Internet, make inferences, and then compile research results.
Introducing deep research | OpenAI
OpenAI unveils a new ChatGPT agent for 'deep research' | TechCrunch
https://techcrunch.com/2025/02/02/openai-unveils-a-new-chatgpt-agent-for-deep-research/
OpenAI says that a major challenge for ChatGPT is its lack of access to the Internet, which has meant that the model cannot access much of the information we use in our daily lives.
The feature to solve this problem is 'Deep research'. Deep research can perform multi-step searches on the Internet and has three main functions: 'content discovery', 'content integration', and 'information reasoning'. In addition, it has the ability to dynamically adapt and update its search plan every time it discovers new information.
Deep research was developed through end-to-end reinforcement learning based on the o3 inference model. This allows the model to plan and execute a multi-step search process and correct course in response to real-time information. In addition, it is possible to view files uploaded by users, perform calculations and create images and plots using Python tools, and embed images from websites.
Regarding the performance of the model, Deep research achieved an accuracy of 26.6% on the benchmark 'Humanity's Last Exam' published by the Center for AI Safety and Scale AI. It significantly outperformed GPT-4o (3.3%), Grok-2 (3.8%), Claude 3.5 Sonnet (4.3%), Gemini Thinking (6.2%), OpenAI o1 (9.1%), DeepSeek-R1 (9.4%), OpenAI o3-mini medium (10.5%), and OpenAI o3-mini high (13.0%). OpenAI emphasizes that the problem-solving process of Deep research is very similar to the thought process of humans who utilize online research.
In addition, in the public benchmark 'GAIA' that evaluates AI for real-world problems, Deep Research achieved new records in tasks requiring reasoning ability, multimodal capabilities, web browsing, and tool proficiency. Specific figures include 74.29% (pass@1) and 78.66% (cons@64) in Level 1, 69.06% (pass@1) and 73.21% (cons@64) in Level 2, and 47.6% (pass@1) and 58.03% (cons@64) in Level 3, achieving an average of 67.36% and 72.57%, respectively. This is an improvement over the previous best of 63.64%.
OpenAI cites that an innovative feature of Deep Research is that it completely removes the response time constraints of conventional models. While conventional models return relatively fast responses, Deep Research can take as long as 5 to 30 minutes to process. OpenAI sees this long processing time as a positive step in the realization of autonomous task execution and as an important step in the roadmap for the development of general artificial intelligence (AGI).
OpenAI evaluates that the output of deep research will be equivalent to a comprehensive, fully cited research paper written by an analyst or subject matter expert. Its uses are diverse, ranging from knowledge work-related tasks to tasks that require extensive web searches, and it can also be used to search for products that meet specific criteria or create content for presentations, OpenAI says.
You can see the actual features in the following announcement by OpenAI, which was filmed live in Tokyo, Japan.
Instruct ChatGPT to 'output comparative data between developed and developing countries on iOS and Android adoption rates, the percentage of people wanting to learn languages, and changes in smartphone penetration rates over the past few years,' and further request 'output in the form of a report with tables and clear recommendations on the most promising emerging markets.' After entering the prompts, click the 'Deep research' button.
ChatGPT responded with a few clarifying questions to get a better understanding of our requirements before we began, such as how we define smartphone penetration and how we measure interest.
Once the user answers these questions, ChatGPT will begin its investigation, and its reasoning will be displayed in the right column.
OpenAI says, 'Deep Research can be used in a variety of knowledge work fields. It can be used not only in market research but also in academic fields such as physics, computer science, and biology. The research process is displayed in the sidebar, and you can see the model's inference process. The model actually opens web pages, analyzes various content such as images, tables, PDFs, and determines the next research steps based on that information.'
Regarding the release schedule, the service will first be launched with the Pro version at the time of announcement on February 3, 2025, and then the Plus version, Team version, Educational version, and Enterprise version will be sequentially rolled out.
Related Posts:
in Software, Web Service, Video, Posted by log1i_yk