Fulford had contributed to the creation of the self-contained artificial intelligence agent, which independently decides what to publish and what to compile into a thorough report. When Deep Research was first made available privately by OpenAI, Fulford claims that whenever it went down, she was flooded with inquiries from coworkers eager to have it again. Fulford says,” The number of individuals who were DMing me made us very thrilled.”
Since going live on February 2 and going lived to the government, Deep Research has gained the support of many people outside the company.
” Deep Research has written 6 studies so far today,” Line CEO Patrick Collison said on X a few days after the item was released. It is in fact superb. Congratulations to the individuals who created it.
Dean Ball, a colleague at George Mason University with a focus on AI plan, wrote,” Deep Research is the Artificial item that really got a significant portion of the DC policy society to start feeling the AG I.”
As part of the ChatGPT Pro plan, which costs$ 200 per month, Deep Research is accessible. It gathers information from a comment, such as” Write me a statement on the Massachusetts health insurance business” or” Tell me about WIRED’s protection of the Department of Government Efficiency,” and then formulates a strategy, searches for appropriate websites, looks through their information, decides what to visit and what information needs to be investigated further, and then decides what to click. It then synthesizes its findings into a thorough record, which may include quotes, data, and charts after exploring for a few hours, often for tens of moments.
Many of the tools that are already marketed as AI agents are basically chatbots connected to basic programs of low level sophistication. The Deep Research concept itself goes through a form of artificial logic before coming up with a strategy and moving ahead with each action. In a area window, the design details this research’s motivations.
Josh Tobin, another OpenAI researcher involved in developing Deep Research, says,” Sometimes, it feels like I need to return, this doesn’t seem that encouraging.” It’s “pretty amazing to learn some of those trajectories, only to understand how the model is thinking.”
Deep Research is seen by OpenAI as a device that could be used for more business tasks, it seems. According to Tobin,” This is a point that we can size,” and that the agent may be trained to carry out certain white-collar tasks. For example, a representative who has access to a company’s internal information was immediately create a statement or display. Tobin states that the longer aim is to “build an agent that is great at many other types of tasks as well as creating information through searching the web.”
Tobin claims that his group was surprised to see how many people were using it to produce code because Deep Research was trained to evaluate and synthesize human-written words. It’s an intriguing string to pull, he claims. We’re never entirely sure what to make of it, they say.
Tobin acknowledges, however, that the device also has significant drawbacks. He claims that it may struggle to distinguish between reliable information and stories. Its trust calibration is weak right now, despite frequently expressing doubt with accuracy.
Age of Reasoning
Ethan Mollick, a teacher at the Wharton School of the University of Pennsylvania who studies business implementation of AI, says that profound study shows how more-capable AI models may manage white-collar work.
Mollick, a frequent user of Deep Research, claims that despite the device being flawed and being most effective when used by professionals who can verify its accuracy, it has impressed professionals he has spoken to. It’s not that it’s perfect or that it beats the best people, Mollick says for senior-level persons. It can do 40 days of medium-level job, Mollick says, and it only takes an hour to check.
It’s unclear whether firms will use these tools to improve their workforce or just change them entirely. That’s what worries me the most, Mollick says.
The possibility of selling tools that can simplify a lot of highly competent office work may explain why OpenAI is considering offering superior agents for a high price. According to a recent report from The Information, the company has informed investors that agents who can perform” PhD-level work” could eventually earn$ 20, 000 per month, although details of such a plan are still undetermined. The document is described as “purely speculation,” according to OpenAI spokesman Kayla Wood.
Deep Analysis shows how border AI study is increasingly focused on both agents and so-called logic models that crack problems down into constituent parts in order to better read and fix them, in addition to hints at changes in white-collar work.
All of OpenAI’s main rivals are creating own reasoning models as well as tools comparable to Deep Research. On December 10, 2024, Google DeepMind released a web research agent with the same name as OpenAI’s tool. Similar features are provided by Elon Musk’s Grok.
Deep Research appears to be the most sophisticated offering right now, in part because it is based on OpenAI’s most advanced reasoning model, called OpenAI o3. Deep Research decides its next steps using a form of simulated reasoning, similar to a traditional large language model’s decision to generate text in response to a query. Although getting models to act without making mistakes is still a challenge, these “agentic” abilities are widely accepted as the next evolutionary step in AI.
Ruslan Salakhutdinov, a computer scientist at Carnegie Mellon University who is also working on web agents, says that” Deep Research is a natural extension of these reasoning models.” However, Salakhutdinov points out that AI agents are still in their early stages, are still prone to errors, and that there will likely be a lot of experimentation and innovation.
To assist in the training of Deep Research, OpenAI hired graduate students and other highly skilled professionals. These users provide training for a reinforcement learning algorithm that enables the model to learn to become a better research assistant by asking queries and then correcting errors.
Alexander Zerkle, a graduate student in microbiology at UC San Diego and the author of Deep Research, claims that “my grandpa is a mathematician.” He desired to demonstrate what is known as the Schroeder-Bernstein theorem. That was given to me by Deep Research, and it spit out a very long proof. Although I don’t understand any of it, he finds it very exciting as a mathematician.
As the popularity of tools like Deep Research increases, they may start to affect how many people use the web, even as the mania that accompanied the chatbot boom fades.
No matter how clever a chatbot is, a model that goes beyond generating text by taking actions and doing valuable work is a different proposition, according to Amelia Glaese, who leads work on alignment at OpenAI. She says,” You have a model that has this very big utility,” and it has learned how to carry out some of the manual research tasks. Then, I believe there is a whole new generation of people who are like,” Wow, this is really useful.”
What are your thoughts on Deep Research and AI agents? Are there any tasks you’d like to see them accomplish? In the comments section below, please tell me everything.
