Models Are Getting Dumber on Purpose

AI models are increasingly being optimized for reasoning and mathematical problem-solving while deliberately sacrificing factual knowledge retention, with smaller models now achieving impressive reasoning benchmarks through parameter efficiency. This trade-off is intentional because reasoning procedures compress better than facts into model weights and don't become outdated, while factual knowledge has a limited shelf life and requires expensive retraining to refresh. Small models like Qwen3.5 9B now outperform larger predecessors on math and logic tasks despite hallucinating facts up to 82% of the time on knowledge-based questions.