Transcription
Hello, AI enthusiasts!
Today, I have some truly incredible news for you that will completely change the world of technology. I say this absolutely seriously because OpenAI has just introduced something that will fundamentally revolutionize everything. This is not an exaggeration; it is pure reality, and you will understand why soon.
One of the limitations, however, of these models is that they don't have access to tools. One of the really core missing tools is the ability to browse the Internet. What this means is that a lot of the things we use in everyday life are now not accessible to them.
So, to announce next, we are introducing Deep Research. We now have a personal research assistant that searches the entire internet—not just superficially, but really delves deep into the material. This assistant truly understands what it reads and even recognizes complex relationships that sometimes even experienced experts miss.
And the most amazing part? This assistant works for you day and night, analyzing thousands of documents in different languages, understanding complex scientific research, and all of this at a speed that would be impossible for humans. This is what makes the new OpenAI technology called Deep Research possible.
As the head of research at OpenAI, Mac, expresses, OpenAI is interested in agents because they believe they will transform knowledge work. They will help companies optimize their processes and make employees more productive. But this will also be really important for consumers.
What sets this apart from all previous AI systems is its ability to truly dive into the material. As Mac explains, they consciously removed time constraints from the model. Typically, models return their answers relatively quickly, but Deep Research models can take time—sometimes even 30 minutes—before returning an answer. This is not a bug; it’s a feature because they believe it’s important for their models to start working on tasks autonomously and without supervision for extended periods.
This is also an important part of their roadmap, as their ultimate goal is a model capable of independently discovering new knowledge. The first step towards this is a model that can find, synthesize, and understand information on the internet.
Let’s dive into the details right away, but first, an interesting fact: the OpenAI team made this announcement straight from Tokyo, where they are currently working with important partners. This already shows how significant this development is, as OpenAI usually makes such announcements from their headquarters in San Francisco.
Interestingly, it is based on the O1 model introduced last year. This was the first model in their series of reasoning models. These models differ from traditional ones in that they think longer before giving an answer, and the longer they think, the better the answer usually becomes.
One of the limitations of these models was that they did not have access to tools, and one of the really important missing tools was the ability to browse the internet. Deep Research fundamentally changes this.
As one of the developers, Isa, explains, Deep Research operates on an optimized version of their soon-to-be-released reasoning model O3. They trained it using reinforcement learning on complex search and other reasoning tasks. Thanks to this training, the model learned to plan and execute multi-step processes while responding to new information in real-time. It can even go back and change its plan if necessary.
The model can not only browse files uploaded by users but also use Python tools for calculations and creating images and diagrams. It can even embed these diagrams in its final answer, just like images from websites. When it cites sources, it cites specific sentences and excerpts.
The system is so advanced that when information is contradictory, it purposefully seeks clarification, which makes this system truly unique. Its ability to understand and process various types of data is remarkable. It can analyze text, understand complex diagrams, interpret technical drawings, process mathematical formulas, and even evaluate audio transcripts.
At the same time, it uses the latest developments in artificial intelligence to connect these different types of information and uncover entirely new relationships. The resulting model can solve problems that would take humans many hours and are very complex.
As Isa explains in the presentation, it also achieves new record values in a number of public and internal assessments. In the Humanity Last Exam, recently released by the Center for Safety AI and Scale AI, which tests the model's abilities across various expert topics, the model achieves a new record of 26.6% accuracy. This test consists of about 3,000 multiple-choice questions and short answers across approximately 100 different subjects.
And what’s really cool is that when you see the thought processes and trajectories of the model, it closely resembles how a human would solve a problem. For example, if I had a really complex problem, Isa explains, I would also start by searching the internet to help myself find a solution. And that’s exactly what it does.
They have seen examples where the model searched for equations in a scientific article for complex physical calculations or in a poetry task where the model needed to determine a very specific meter for a new poem. It searched for examples of other existing poems to learn how to arrive at the answer.
But let’s look at a specific example straight from the presentation. One of the product managers at OpenAI wanted to find out if it makes sense to develop a new language learning app. He asked to analyze the usage statistics for iOS and Android, the percentage of people wanting to learn a new language, and the changes in mobile penetration over the past few years, with a distinction between leading developing countries and emerging markets.
He wanted this information in a formatted report with tables and a clear recommendation on the best promising opportunities. This is a task that would typically take hours, but with Deep Research, he could start it immediately. The system asked clarifying questions to clarify the requirements, just like a real analyst would.
After all, if Deep Research is exploring for 5-30 minutes, you need to ensure that the requirements are correct from the very beginning. In just 11 minutes, the system thoroughly analyzed 29 different sources and created a perfectly formatted report with a detailed introduction, various adoption trends, and a wealth of different data.
But it gets even more interesting because OpenAI is already planning the next steps for development. As Mark explains, what we are launching today is just the surface of what can be imagined with Deep Research.
Today, we have the Deep Research agent that searches the internet, but one can imagine that the same agent could connect to user context or even to corporate data repositories. Deep Research is important for their AGI roadmap. They believe in agents that work autonomously for longer and longer periods to solve very complex tasks.
They believe that the ability to work on a task for 30 minutes motivates much larger investments in computing. Sam Altman, the CEO of OpenAI, even said that it can already take on a significant percentage of all economically valuable tasks.
Imagine what that means! We are talking about a system that can independently conduct complex research, analyze data, and uncover new relationships—a system that does not just summarize existing knowledge but actively generates new knowledge.
This is a giant leap towards true artificial intelligence, and the most exciting part is that we are just at the beginning of the journey. The technology will continue to improve, develop new capabilities, and tackle increasingly complex tasks.
We are here, live, witnessing a new chapter being written in the history of technology, and I can’t wait to show you in the next videos what is possible with Deep Research.
If you enjoyed the video, leave a like and subscribe to the channel. And if you activate the bell, you will even receive notifications when we upload a new video.
I hope to see you next time. Until then, stay real!