In its first major release since the turbulence was caused by Deeppeek, Openai notably named a new AI tool called Deep Research on Sunday, February 2nd.
Simply put, Deep Research is an AI tool that allows you to collect information from across the Internet and synthesize it to produce comprehensive reports at the level of a research analyst.
Openai announced its deep search capabilities in a demo video posted to YouTube. Last week, he personally introduced the AI tools to US lawmakers and other government officials in Washington, D.C., according to the New York Times.
The AI startup behind CHATGPT is placed in an unstable position by DeepSeek. This claims that its inference model (R1) is comparable to OpenAI’s O1 model, using a lower advanced graphics processing unit (GPU), which significantly reduces costs.
This breakthrough in computational efficiency led to the perception that Openai is losing its position in a Chinese startup like AI Arms Race’s Deepseek. In response to the talks around Deepseek, Openai CEO Sam Altman said the company would “provide a much better model” and “uplift some releases.”
The Deepseek R1 is an impressive model, especially focusing on what can be offered at a price.
We clearly offer a much better model and it’s legal to have a new competitor! We’ll raise some releases.
– Sam Altman (@sama) January 28, 2025
Deep Research is the second AI agent Openai released in a month. A few weeks ago, we previewed an AI agent called an AI agent that allows you to search the internet, visit websites on your behalf, shop for groceries, and make restaurant reservations.
What is deep research? How was it developed?
Openai’s operator agent is built on the multimodal large language model (LLM) GPT-4O, but Deep Research is the first to feature a version of the upcoming O3 model optimized for web browsing and data. It is thought to be an AI tool. analysis.
According to Openai, Deep Research performs complex tasks that require “a wide range of context and information gathering from diverse online sources” based on O1’s inference capabilities.
The story continues under this ad
In-depth research has been trained on real-world tasks across a wide range of domains that require the use of browsers and Python tools. The training method employed was end-to-end reinforcement learning.
“Through that training, I learned to plan and execute a multi-step trajectory to find the data I need for real-time information when needed,” he said in a blog post published Sunday.
How do you use deep research? What can it do?
In response to user prompts, Deep Research provides comprehensive reporting after finding, analyzing and compositing content such as text, images, PDFs and more from several online sources.
AI-generated research reports include composite graphs and images from the website, along with clear citations and summary of their thinking. Users can also upload specific files and spreadsheets to add context to the prompt.
The story continues under this ad
Openai argued that deep research is particularly good at discovering hard-to-find information, which usually requires searching multiple websites. For example, users can describe scenes from any TV show, delve into the web, and ask them to find the exact episode in which it happened.
“(Deep research) can perform complex research tasks that can take people from 30 to 30 days,” says Kevin Weil, Chief Product Officer at Openai, quoted. It was done.
However, the company admits that AI tools may still be wrong. “It may be difficult to distinguish authoritative information from rumors, and it is currently showing weakness in calibrating confidence, often showing inability to accurately communicate uncertainty.” Openai said.
Currently, Deep Research is accessible via the web version of ChatGPT as part of the Pro Subscription package. Users must select the (Deep Search) tab in Message Composer before entering the prompt.
The story continues under this ad
“When you start running, you will see a sidebar and a summary of the sources used (…) you will receive a notification when the investigation is complete. The final output will arrive as a report in the chat.” The company said.
In-depth research is also very computationally intensive. In other words, the longer it takes to investigate a query, the more processing power it will need. Currently it is not accessible in the UK and the European Union (EU) regions.
How did Deep Research work in benchmark tests?
Openai claimed that the optimized version of O3 Powering Deep Research achieved a new score of 26.6% accuracy in the final test of humanity. This is a recently released benchmark test that tests AI models in a wide range of subjects with expert-level questions.
“Compared to Openai O1, the greatest benefits (of deep research) have appeared in chemistry, the humanities, social sciences and mathematics.”
The story continues under this ad
Furthermore, the O3 model, which is the basis of Deep Research, is another Gaia that evaluates AI models for inference, multimodal flow encyability, web browsing, and tool use capabilities, with questions spanning three levels of difficulty. Public benchmark tests have been put on top.
Openai then said in its AI-generated research report, deep research will be upgraded to embed images, data visualizations, and other analytical outputs. At the end of the user, more specialized data sources can be added to “more robust and personalized” the output of deep research.
