OpenAI has unveiled a new family of AI models called OpenAI o1, designed to spend more time thinking before responding to user queries.
The company explained, in an official statement, that the new family will have the ability to solve more complex tasks and more difficult problems in the fields of science, programming and mathematics, compared to previous models.

The American company explained that its new models specialize in dealing with complex thinking and reasoning tasks, and this is "a whole new level of capabilities for artificial intelligence models", so OpenAI decided to change the way it names future models, and unveiled the first two models within the new family, OpenAI o1 and OpenAI o1-mini. OpenAI added: "We are also continuing to develop and launch new models under the umbrella of the GPT model family."

How do the new OpenAI o1 models work?

OpenAI trained the new model family to take longer to think about solving the dilemmas presented by the user, before providing a response to them, closer to the human way of thinking During the training, the ability of OpenAI o1 AI models to improve the thinking process, use different strategies, and recognize their mistakes, to improve the methodology of thinking in the future, developed.

OpenAI o1 model capabilities 

The company pointed out that during the tests, the new generation of smart models succeeded in proving their ability to think with performance similar to the way students of a scientific doctorate think, in solving tasks in the fields of physics, chemistry and biological sciences.

In the IMO (International Olympiad in Mathematics), GPT-4o solved only 13% of all mathematical problems, while OpenAI o1 models reached 83% of the correct solutions to test problems.

However, the new models have proven not to have the same advantages as previous OpenAI models, for example, GPT-4o offers superior advantages in browsing the Internet for information and analyzing files attached to user commands from different images and document types, but the new OpenAI o1 models do not offer this level of diversity in performance, so in most aspects of using ChatGPT, the GPT-4o family will be the most able to meet the needs of users in the near term.

OpenAI explained that it is currently considering adding web browsing, file analysis and other features to the OpenAI o1 models, to extend their use to all users.
Besides OpenAI o1, the company also introduced a "mini" version of the same model, which is mainly aimed at developers, and offers greater speed in performance as well as a lower cost of 80% than the basic version.
To ensure that its new models offer a level of security that is difficult to overcome, OpenAI has developed a new approach in safety training, based on exploiting the cognitive capabilities of the new models, to make them adhere to the company's security standards and guidelines, and apply them more effectively.

One of the company's methods to test the ability of new smart models to adhere to security rules, in case a user tries to bypass them and circumvent them, focused on what is known as jailbreaking, where the OpenAI o1 model achieved an average of 84 out of 100 degrees, while the GPT-4o model achieved only 22 degrees. The company also conducted a full update of its internal security and safety policies, internal governance, as well as its own testing frameworks, in cooperation with a number of artificial intelligence safety institutes in the United States and Britain, as safety testing teams at those institutes obtained early access to a research version of the new models for testing.

OpenAI stressed that the new models and their groundbreaking cognitive abilities will be useful for science, mathematics and programming professionals, for example, it will be of great help to researchers in the healthcare sector in explaining cell sequence data, it will help physicists create complex mathematical formulas required in the field of quantum optics, and it will facilitate the task of developers in building and implementing multi-step workflows.

The new OpenAI o1 models, starting September 12, are available paid to subscribers of ChatGPT Plus and ChatGPT Team services, so the new models, o1 and o1 mini, can be manually selected from the main platform user interface.

The company will provide an average usage of 30 messages for OpenAI o1 per week and 50 messages for OpenAI o1-mini per week, and OpenAI is working to raise those limits, as well as enabling the ChatGPT platform itself to automatically switch between available AI models as ordered by the user. The company also plans to offer the OpenAI o1-mini model to the general public of ChatGPT users for free soon. 

ASharq News