OpenAI said that o1 and o1-mini AI models can apply different strategies and learn from mistakes to solve science, coding, and maths problems