OpenAI Unveils New ChatGPT Model With Nearly Human Logic

OpenAI has presented o3, a completely new breakthrough ChatGPT model. According to tests, the neural network has made another step towards hypothetical self-learning AI (AGI).

The new model is a direct successor to o1, but the number «2» in the name had to be abandoned due to a possible conflict with the British provider O2. Direct access to the neural network has not yet been opened, but researchers can already apply for testing. A preliminary version of o3 will appear soon, promises OpenAI founder Sam Altman. The official launch is expected in January 2025.

The new version of AI differs in that it carefully checks the facts before responding, so its response may be slower than other models. In addition, it uses a chain of thoughts before responding to the user. The model can analyze the task and plan ahead, performing a series of actions over a long period to find a solution. In fact, the AI ​​reasons almost at the level of a human, although errors and «hallucinations» are not excluded.

Experts have already stated that o3 has come close to self-learning AI, which, according to the head of OpenAI, will appear in the coming years. As ARC-AGI co-author François Chollet stated, the model scored 87.5% in a test that reveals the ability to acquire new skills without preliminary training on data. However, this indicator is not entirely correct — the neural network copes well with complex tasks, but simple ones cause problems.

In the near future, independent researchers will publish the results of experiments demonstrating the capabilities and risks of o3. One of these tests, for example, revealed the ability of ChatGPT o1 to consciously lie.