Video

OpenAI has unveiled GPT-6 Astra. Delivering performance that significantly surpasses previous models in evaluations such as ARC-AGI, it performs tasks ranging from game creation and 3D modeling to circuit design and direct computer operation, even earning evaluations as an AGI-level model. However, along with these performance improvements, a new problem has emerged. Astra often leaves shorter natural language chains of thought (CoT) or none at all compared to previous models, making it more difficult to monitor abnormal model behavior through the chain of thought alone. In this OGraph video, we examine the story of GPT-6 Astra along with five graphs on how to control an unfamiliar intelligence that humans are finding increasingly difficult to understand.

(Reported by An Hyemin | Filmed by Hwang Se-hoe and Kim Sang-yun | Edited by Lee Ki-eun | Designed by Ahn Jun-seok | The Intern: Shin Yeon-seong | Produced by Intellectual Content IP Team)