Sam Altman and other executives at OpenAI are discussing the performance of the new o3 model on the ARC-AGI test. The latest release from OpenAI, o3, was revealed through a video infomercial on Dec. 20. Although the model has not been released to the public yet, there are early indications of its capabilities. One outsider had the opportunity to test o3 on the ARC-AGI challenge, where it scored 76% accuracy, surpassing human scores for the first time. This achievement is considered a significant step forward in AI capabilities, with the potential to compete with human work in the near future. The test, designed to measure adaptability to novelty, marks a qualitative shift in AI capabilities. Despite these advancements, the exact methods used by o3 to solve the challenges are still unclear due to OpenAI’s closed-source model. The model’s architecture and approach to solving tasks are speculated to be different from previous models, incorporating a significant amount of test-time search. Despite some uncertainties, o3’s performance on the ARC-AGI test indicates a promising future for AI capabilities.
