Openai has announced a new AI “Agent” designed to make detailed and complex research, using Chatgpt, a chat bot platform equipped with AI.
Appropriately, it is called deep -search.
Openai is a blog post published on Sunday, and these new features are designed for those who “focus on finance, science, policies, engineering, etc., for those who conduct intensive knowledge work, and are thoroughly and accurate. You need a reliable research. ” The company also added that it may be useful for those who create a “purchase that requires a cautious survey of cars, appliances, and furniture”.
Basically, Chatgpt Deep Research not only requires simple answers and summaries, but also for instances that need to be eager to examine information from multiple websites and other sources.
Openai has stated that today’s Chatgpt Pro users have restricted 100 queries a month, so that PLUS and team users support the next, so that deep surveys are used. (Openai has targeted plus rollout about a month later, and the paid users’ queries will soon be “very high”.) This is the launch of Geo Targets. is. Openai had no release timeline shared with Chatgpt customers in the UK, Switzerland, and European economy.
To use Chatgpt Deep Research, simply use the option to attach a file or spreadsheet to select the composer “deep -search” and then enter the query. (This is now an experience -only experience, and the integration of mobile and desktop apps will appear in the second half of this month.) A deep survey can take 5-30 minutes to answer questions. 。
Currently, the output of Chatgpt Deep Research is only text. However, Openai stated that embed images, data visualization, and other “analysis” outputs will be added immediately. The roadmap also has a function to connect “more specialized data sources”, such as “subscription base” and internal resources.
The big problem is how accurate Chatgpt Deep Research. After all, AI is incomplete. In the “deep research” scenario, the hallucinations that are particularly harmful and other types of errors are likely to occur. Perhaps the reason is that Openai is completely documented all Chatgpt’s deep research output, and it can be easily referenced and verified by a clear quotation and () thinking.
Judges have shown whether these easing is sufficient to fight AI mistakes. The web search function equipped with AI of Chatgpt and Chatgpt search is rarely performed frequently and gives the wrong answer to questions. TechCrunch tests showed that Chatgpt search had less usefulness than Google searching for specific queries.
In order to enhance the accuracy of Deep Research, Openai uses a special version of the recently announced O3 “River” AI model, and is “Real World Tasks that require browser and Python tools”. Trained through reinforcement learning. Enhanced learning essentially “teach” models through trial and error, achieving specific goals. As the model approaches the goal, ideally, we receive a “reward” that improves future tasks.
This version of the Openai O3 model states that it is “optimized for web browsing and data analysis” and “needs a lot of texts, images, and pdfs on the Internet, and needs inference for interpreting and analyzing. Pivot in response to the information source to the encounter.
The company stated that he had tested Chatgpt Deep Research using the last test of mankind. This is an evaluation that includes more than 3,000 expert -level questions in various academic fields. The O3 model, which supplies the power of Deep Research, has achieved 26.6 % accuracy, but it may seem to fail, but the last exam of mankind goes ahead of the model. It was designed to be stricter than other benchmarks. According to Openai, Deep Research O3 models are prior to Gemini Thinking (6.2 %), GROK-2 (3.8 %), and Openai’s unique GPT-4O (3.3 %).
Nevertheless, Openai states that Chatgpt Deep Research has a limit, making mistakes and doing incorrect inference. The company said that deep research could have difficulty distinguishing prestigious information from rumors, and often cannot tell it when it is uncertain.
For those who are worried about the effects of the generated AI on the students, or those who are trying to find information online, the output that is appropriately quoted in this type is an outline of a seemingly simple chatbot. It sounds more attractive. However, make sure that most users actually lead the output to actual analysis and double checks, or simply handle it as a more professional textbook.
And if this sounded all familiar, Google actually announced the same AI function with exactly the same name two months ago.
