The 12 Days of OpenAI Conclude with the Launch of a New Model for the Coming Year.
The launch of o3 and o3-mini is being prepared.
OpenAI has recently introduced its new artificial intelligence models, o3 and o3-mini, which focus on improving reasoning capabilities compared to previous models, o1 and o1-mini. These advancements were part of the final day of a special event celebrating OpenAI, where Sam Altman, the CEO of the company, was present to unveil these innovations.
The models o3 and o3-mini have been designed to take time to reflect before issuing a response, allowing them to reason better in their answers. While the mini version is intended for specific and limited tasks, both versions aim to take a significant step towards artificial general intelligence (AGI), although the company acknowledges that they are still far from achieving this goal.
One of the standout features of o3 is its ability to take time to process information before responding, thanks to a system that allows for a "private thought chain." This helps avoid common AI mistakes, such as providing incorrect answers with excessive confidence. Though this approach may slightly slow down responses, it results in greater accuracy in critical areas such as mathematics, science, and programming.
Additionally, users will be able to adjust the model's reflection time by selecting from settings that allow for quick responses or, conversely, more thorough reasoning. During testing, o3 has shown to significantly outperform its predecessor.
However, o3 does not represent complete general intelligence, as it cannot replace humans in all tasks nor does it meet the definition of AGI established by OpenAI, which refers to the ability of models to surpass humans in high-value economic projects. This limitation also raises questions about OpenAI's partnership with Microsoft, as the achievement of AGI could alter their exclusive access agreement to the most advanced models.
In terms of availability, o3 and its mini version are currently not accessible to the general public. OpenAI is allowing security researchers to explore these innovations through Copilot Labs, and o3-mini is expected to be available by the end of January, followed by the full o3 model. This controlled launch reflects the complexity and potential of these new systems.
Thus, o3 is shaping up to be an indication of the direction artificial intelligence is taking, presenting tools that not only generate content but also process problems more reflectively, marking a new horizon in this field. The coming months will be crucial to assess whether o3 meets the generated expectations.