OpenAI unveils a new ChatGPT agent for ‘deep research’

OpenAi announces a new AI “agent” that is designed to help people to carry out in-depth, complex research using chatgpt, the AI-driven chatbot platform of the company.

The deep investigation is called in a suitable way.

Said openi A blog post Published Sunday that this new capacity is designed for “people who do intensive knowledge in areas such as finance, science, policy and engineering and need thorough, accurate and reliable research.” It can also be useful, the company added, for anyone who “makes purchases that usually require carefully research, such as cars, devices and furniture.”

In short, chatgpt deep research is intended for cases in which you not only want a quick answer or summary, but instead must consider diligent information from multiple websites and other sources.

OpenAi said that today makes in -depth research available for Chatgpt Pro users, limited to 100 questions per month, with support for plus and team users who come afterwards, followed by Enterprise. (OpenAI will focus on a plus outlet in about a month in about a month, the company said, and the query limites for paid users must soon be “considerably higher”.) It will be a geo-oriented launch; OpenAi did not have a release time line to share for chatgpt customers in the UK, Switzerland and the European Economic Area.

To use chatgpt deep research, simply select “Deep Research” in the composer and then enter a query, with the option to add files or spreadsheets. (For the time being it is an experience with the web, where the integration of mobile and desktop app will arrive later this month.) Deep research can then take 5 to 30 minutes to answer the question, and you will receive a notification when the search is completed.

Currently, Chatgpt Deep Research’s output is only text. But OpenAi said it is planning to add enclosed images, data visualisations and other “analytical” outputs. The route map also has the possibility to connect ‘more specialized data sources’, including ‘subscription -based’ and internal sources, added.

The big question is, how exactly is chatgpt deep research? After all, AI is imperfect. It is susceptible to hallucinations and other types of errors that can be particularly harmful in a “deep research” scenario. That is perhaps the reason why OpenAi said that every chatgpt deep research output “will be fully documented, with clear quotes and a summary of [the] Thinking, making it easy to refer and verify the information. “

The jury will assume whether those mitigations will be sufficient to combat AI errors. OpenAI’s AI-driven web search function in chatgpt, chatgpt search, not infrequently makes blunders and gives wrong answers to questions. Testing TechCrunch discovered that ChatGPT search produces less useful results than Google search for certain questions.

To strengthen the accuracy of deep research, OpenAI uses a special version of the recently announced O3 “reasoning” AI model that has been trained by strengthening learning to reinforce “Real-World tasks required for browser and python tool use” . In essence, reinforcement learn a model through trial and error to achieve a specific goal. As the model gets closer to the goal, the virtual ‘rewards’ receives that ideally make better in the task in the future.

It said that this version of the OpenAI O3 model “is optimized for web browses and data analysis”, adding that “it uses the reasoning to search, interpret and analyze huge amounts of text, images and PDFS on the internet if necessary in response to information it encounters […] The model is also able to browse, plot and itteries on graphs using the Python tool, both generated graphs and images of websites in its answers and to enclose specific sentences or passages from the sources to quote. “

The company said the chatgpt tested deep research with the help of The last exam of humanityAn evaluation that includes more than 3,000 questions at expert level in various academic fields. The O3 model that flowed deep investigation achieved an accuracy of 26.6%, which may resemble a failing figure – but the last exam of humanity was designed to be more difficult than other benchmarks to stay ahead of the progress of The model. According to OpenAI, the deep research O3 model was far ahead of Gemini Thinking (6.2%), Grok-2 (3.8%) and OpenAi’s own GPT-4O (3.3%).

Nevertheless, OpenAI notes that chatgpt has deep research limitations, sometimes making mistakes and incorrect conclusions. Deep research may have difficulty distinguishing authoritative information from rumors, the company said, and often not conveying when it is uncertain about something – and it can also make formatting errors in reports and quotes.

For everyone who is worried about the impact of generative AI on students, or on anyone trying to find online information, this kind of profound, well-cited output is probably more attractive than a deceptively simple chatbot companion without quotes. But we will see if most users will actually subject the output to real analysis and double control, or whether they simply treat it as a more professional -looking text to copy.

And if all this sounds familiar, Google actually announced a similar AI function with exactly the same name less than two months ago.