OpenAI researcher Dan Roberts at AI Ascent 2025 reveals how reasoning capabilities in AI are evolving, why test-time compute is revolutionizing the field, and the controversial shift to prioritizing reinforcement learning over pre-training.