OpenAI launches o1, its first AI model with reasoning capabilities to solve more complex tasks and problems

September 13 (Portaltic/EP) –

OpenAI has launched o1, the company’s first Artificial Intelligence (AI) model with reasoning abilitieswhich spends “more time thinking before responding” to solve complex tasks and more difficult problems than its predecessor models, in the subjects of science, programming and mathematics.

The technology company led by Sam Altman continues to advance in AI, in this case, with the development of a new series of models designed to get them to spend more time thinking about problems before respondingso that they act more similarly to how a person would reason.

In this sense, OpenAI has launched the first model of this new series with advanced reasoning capabilities, referred to as o1, and is now available through ChatGPT and the company’s API, in preview format.

This is a model that can reasoning about complex tasks and therefore can solve problems of greater difficulty than previously released models in different areas, whether science, programming or mathematics, as the technology company has stressed in a statement on his blog.

To achieve these capabilities, the company has noted that it has trained the models to act slower, That is, to make them spend more time thinking before offering a response. Thus, through this training, the models learn to refine your thought processas well as to try different strategies to solve a problem, in order to choose the best answer, acknowledge their mistakes.

As shared by the OpenAI Research Director Jerry Tworek, in statements to The Vergethe training for the o1 model has been different from previous models. As he explained, it has been trained using “a completely new optimization algorithm and a training data set specifically designed for it.”

This translates into a training method that, through a system of rewards and penalties, Teaches the model to think for itselfrather than imitating data patterns. In this way, Tworek has indicated that they have made the model “less mind-blowing” in his responses, although he has acknowledged that the problem has not been completely resolved.

As a result of this training, OpenAI has noted that this new model performs similarly to how PhD students work on challenging benchmark tasks in physics, chemistry and biology. The same is true in areas such as mathematics and coding.

Specifically the o1 series stands out for generate and debug complex code accurately. Thus, as they were able to verify, when trying to solve a qualifying exam for the International Mathematical Olympiad (IMO), the new o1 model obtained a score of 84, while the previous model, GPT-4opresented in May of this year, obtained a score of 22.

With all this, OpenAI has indicated that the o1 model can be useful for solving complex problems in science, coding, math and similar fields. It can also be used to carry out research on the health sector or physicsfor example, to generate complicated mathematical formulas needed for quantum optics.

Following this line, given the new reasoning capabilities offered by this model, OpenAI has pointed out that they have also reinforced security. For what they have carried out “rigorous assessments” using their Readiness Frameworkthat is, work teams that carry out review processes at the board level, including its Safety and Security Committee.

OPENAI O1-MINI

In addition to the o1 model, OpenAI has also launched the version o1-mini, which is a model of reasoning faster and cheaper, “particularly effective in coding,” as the company describes it.

Being a model with a smaller size, it is a 80 percent cheaper than OpenAI o1 previewHowever, it remains a “powerful and cost-effective” model for performing tasks that require reasoning but not necessarily “broad knowledge of the world.”

To use OpenAI o1 and OpenAI o1-mini, users must have ChatGPT Plus and Team. Thus, it will be enough to choose one of these new options from the model selector to begin testing the reasoning capabilities. However, the company has indicated that they have frequency limits weekly for use. Specifically, 30 messages for o1 preview and 50 messages for o1-mini.

For their part, users of ChatGPT Enterprise and Edu will have access to both models from the next weekAs for developers, those with API usage level 5 can now start prototyping with a speed limit of 20 RPM. Finally, OpenAI has announced that it will offer access to o1-mini for all users of ChatGPT Free in the future.