Models think on much longer horizons than training suggests

Universal language of thought spans multiple languages internally

Hours of human effort required to understand model circuits

Source: AI Neuroscience · Anthropic Research Team · September 10, 2025