OpenAI made a big announcement on the last day of its 12-day “Shipmas” event, introducing the new o3 model family. This model family is actually a successor to the o1 “reasoning” model released earlier this year. The o3 and its smaller version, the 03-mini, are said to be highly optimized for specific tasks. Details in our news…
OpenAI is almost close to Artificial General Intelligence with its new o3 model!
The company claims that the o3 model approaches Artificial General Intelligence (AGI) in some circumstances. However, this claim has some doubts and reservations for now. The o3 model is slightly different from other artificial intelligences as a “reasoning” model. Here are the prominent features of o3:
- Specialized thought chain: The model simulates the thinking process before completing a task, planning a series of actions and evaluating relevant issues to reach a solution.
- Variable thinking time: Users can tune the model’s performance by choosing low, medium or high levels of computation time (thinking time).
- Self-validation: The model checks its answers internally, leading to more accurate results.
Still, while o3 provides more reliable solutions in areas such as physics, math and science, it takes longer than other models. OpenAI reports that o3 is close to AGI in some tests. For example, in a test called ARC-AGI, o3 achieved 87.5% in the high processing power setting. This measures the ability of an AI system to acquire new skills beyond training data.
Tops benchmark tests
o3 achieved impressive results in different benchmark tests. In the SWE-Bench Verified test, it showed an increase of 22.8%. In the American Mathematics Exam, it achieved 96.7%, missing only one question. It achieved 87.7% on the GPQA Diamond set of graduate-level biology, physics and chemistry questions.
