OpenAI has just announced its next generation AI model, o3, with impressive capabilities exceeding current models in coding, math, and reasoning benchmarks. o3 achieves a 71.7% accuracy rate on the Sweet Bench coding benchmark, significantly outperforming its predecessor, o1. It also excels in competition math, achieving a near-perfect score of 96.7% and surpasses human performance on the Arc AGI benchmark with a score of 87.5%. These achievements suggest that o3 is approaching the definition of AGI, as it outperforms humans in most economically valuable work. The announcement also included o3 mini, a more cost-effective version with impressive performance and faster response times. Both models will be available for public safety testing, with researchers invited to apply for early access. OpenAI emphasizes the importance of external safety testing and encourages researchers to explore the models' capabilities and potential risks.
No comments:
Post a Comment