/blog/research/when-ai-overthinks-the-inverse-scaling-problem