OpenAI, creator of ChatGPT, presents new models and believes it is closer to a General AI

OpenAI has presented its new family of models Artificial intelligence (AI), o3 and o3-mini, which offer improved performance over previous models and are closer to “Artificial General Intelligence” (AGI).

The company led by Sam Altman has unveiled the successor models of the o1 family within the framework of its latest announcement within 12 days of OpenAI, which they plan to launch “at the end of January” in the case of the o3-mini model, and later in the case of o3.

Specifically, like the o1 models, the new models in the o3 family continue to expand reasoning capabilities, spending more time thinking before responding to improve their skills when solving complex tasks and problems. Thus, when a request is launched, o 3 analyzes related indications and explains its reasoning before offering a definitive response.

Likewise, these new AI models offer performance improvements compared to their predecessors, for example, according to the results of the American Invitational Mathematics Examination tests, o3 achieves an accuracy score of 96.7%, compared to 83.3 % that obtains o1, as the company has highlighted in an explanatory video on YouTube.

As OpenAI Senior Vice President of Research Mark Chen explained, these performance improvements mean that o3 “often fails on a single question,” demonstrating more accuracy in its answers than previous models.

The same happens in the rest of the usual tests that OpenAI subjects its models to, in which o3 also obtained good results, so the company resorted to new, more challenging tests to put it to the test.

For example, the technology company used the ARC-AGI test, specially developed to measure the capabilities of AI models in comparison with the human mind, with which it managed to achieve a score of 87.5%, a high figure considering that the 85% threshold is comparable to human performance.

In this sense, OpenAI has detailed that o3 represents an advance for Generative Artificial Intelligence, and that, even with its lowest score in the ARC-AGI test, it managed to triple the performance of o1.

MODEL O3-MINI

On the other hand, OpenAI has also presented o3-mini. Specifically, it is based on a smaller and more refined version of the model, optimized to carry out specific tasks.

This model uses the adaptive thinking time feature, recently introduced in the API, to offer three reasoning modes, choosing between low, medium and high. This way, users can adjust the time the AI ​​takes to think about a problem before offering a response. However, it must be taken into account that, if a low reasoning time is chosen, the model responses may suffer more inaccuracies.

Likewise, as OpenAI has shared, the o3-mini model is capable of offering results comparable to the current reasoning model o1, however, it obtains these results with a fraction of the computational cost of the o1 model.

With all this, although the new o3 and o3 mini models have not been officially launched, OpenAI has announced a program for security researchers where they can register to preview the o3-mini. Following this line, Altman has announced that they plan to launch the o3-mini model at the end of January 2025 and, after that, the o 3 model.