AI Beats Olympiads Yet Struggles With Basic School Math

The year 2025 marked a remarkable milestone when Google DeepMind and OpenAI systems delivered gold medal performances at the International Mathematical Olympiad. These AI models solved problems typically cracked only by a select few of the world's brightest young mathematicians. Yet, these same systems frequently stumble over basic arithmetic that any middle school student could handle. This striking paradox reveals something fundamental about contemporary AI: we are witnessing the emergence of a jagged intelligence, where machines exhibit superhuman ability in specific domains while failing at tasks we deem elementary.

The Olympiad Triumph

The International Mathematical Olympiad represents the pinnacle of pre-university mathematical competition. Each year, top students tackle six problems requiring deep insight, creativity, and advanced proof techniques. In 2025, AI from both Google DeepMind and OpenAI scored 35 out of 42 points, securing gold medals. DeepMind’s AlphaGeometry 2 solved a complex geometry problem in just 19 seconds, while AlphaProof tackled number theory and algebra problems that stumped most human participants.

These breakthroughs build on years of steady progress. The systems utilize formal mathematical languages like Lean to construct rigorous proofs and employ methods such as curriculum learning, training on problems of increasing difficulty. This process enables AI to grasp complex relationships between mathematical objects, recognize subtle patterns, and formulate elegant proofs.

The Elementary Struggle

The very AI that triumphs at the Olympiad often falters on seemingly trivial tasks. Asked to multiply large numbers, it may confidently produce an incorrect answer. Its performance on other basic arithmetic operations can be similarly unpredictable. The issue extends beyond simple calculation. These systems often struggle with word problems that require tracking multiple quantities, understanding real-world context, or sequentially applying fundamental operations.

This weakness is rooted in how these models operate. Large language models predict the next piece of text based on patterns in their training data. When they see "2 + 2," they correctly output "4" not because they comprehend addition, but because this sequence appears ubiquitously in their training. Present them with unusual, rarely-seen calculations, and their performance plummets. They are pattern-matching engines, excelling with clear, consistent patterns but struggling when forced to compute something novel.

The Architecture Paradox

The contradiction between Olympiad success and arithmetic failure points to a deeper architectural issue. Modern AI excels at problems solvable through pattern recognition, logical deduction, and systematic exploration of solution spaces. Olympiad problems, though difficult, often possess elegant structures that AI can leverage. The systems can explore proof strategies, verify logical steps, and build upon established mathematical frameworks, operating in a world governed by symbols, rules, and logic.

Basic arithmetic, paradoxically, presents a different challenge. It demands precise manipulation of quantities, not pattern matching. It requires an understanding of numerical magnitude and relationships that cannot be approximated. When an AI model treats arithmetic as a language modeling task, it sees numbers as tokens to predict rather than quantities to compute. This fundamental mismatch between task requirement and model architecture creates the observed performance gap.

Training Data and Its Limitations

AI capabilities are heavily shaped by training data. Mathematical proofs and advanced problems often exist online in well-structured formats like academic papers, textbooks, and educational resources, which provide clear examples of reasoning. The internet abounds with discussions on mathematical concepts and problem-solving strategies, creating a rich corpus for learning advanced thinking.

Elementary mathematics faces a different issue. While basic arithmetic is common online, it's rarely accompanied by detailed explanations of the underlying process. Simple calculations are stated as facts, not explained as procedures. The training data contains the results of computations but not the step-by-step reasoning, creating a gap in understanding that manifests as poor performance on fundamental tasks.

Implications for AI Development

This jagged intelligence has crucial implications for AI design and deployment. Success in complex tasks does not guarantee competence in simpler ones. An AI that proves theorems might fail to balance a checkbook; a system that writes code might struggle with basic counting. This reality demands careful assessment of capabilities and limitations for real-world applications.

The phenomenon also underscores the value of hybrid approaches. Rather than expecting a single model to handle everything, we may need specialized systems for different tasks. Combining symbolic computation for arithmetic with language models for reasoning could yield more reliable solutions. The future may lie in orchestrating multiple specialized systems rather than pursuing a single, monolithic intelligence.

The Path Forward

Acknowledging jagged intelligence clarifies the path toward more capable AI. Researchers are developing ways to integrate computational tools into language models, allowing them to offload arithmetic to dedicated calculators. New training strategies teach models when to use external tools instead of internalizing every skill. This mirrors human intelligence, where we use tools for computation and focus our mental energy on higher-level reasoning.

Ultimately, the paradox of jagged intelligence teaches humility. These systems are neither universally superior nor uniformly limited. They possess a complex tapestry of strengths and weaknesses that we must understand to use and improve them effectively. Progress requires not just expanding AI's capabilities but also addressing its fundamental gaps. Machines that prove theorems yet fail at basic addition remind us that intelligence—artificial or human—remains a multifaceted phenomenon defying simple definition.

The Bottom Line

AI's ability to solve Olympiad problems but not simple math demonstrates that intelligence develops unevenly. A system can be brilliant in one area and surprisingly weak in another. Understanding this jagged profile is essential for designing and applying AI responsibly. Instead of a single model for all tasks, the solution may involve combining different approaches that leverage each system's strengths. Real-world progress will come from building AI that works reliably in practice, not from assuming it will excel at everything.