I compared my daughter against SOTA models on math puzzles

Michał Prządka conducted an experiment comparing the mathematical reasoning abilities of advanced AI models to those of his 11-year-old daughter, Anika, using puzzles from the GMIL competition. The results showed that while some AI models performed comparably to Anika, others struggled significantly. Anika solved 7 out of 8 puzzles correctly, while the AI models varied in their success. The benchmark included models like GPT-o3-mini, which achieved a perfect score, and DeepSeek-R1, which also performed well but relied on brute-force methods. In contrast, models like Sonnet-3.5 and Gemini-Flash had lower success rates. The experiment highlights the current capabilities of AI in mathematical reasoning and suggests that while AI can match human performance in some areas, there are still limitations. Prządka expressed pride in his daughter's skills, noting that they are on par with advanced AI models, and he plans to extend this analysis to more challenging puzzles in the future.

- Anika scored 7 out of 8 correct in the math reasoning benchmark.

- GPT-o3-mini achieved a perfect score, while other models had varying success rates.

- The experiment highlights the strengths and weaknesses of AI in mathematical reasoning.

- Prządka plans to explore more difficult puzzles to further assess AI capabilities.

- The results indicate that AI can match human performance in certain tasks but still has limitations.