When AI models achieve superhuman performance on specific benchmarks like coding challenges, it doesn't solve real-world problems. This is because we implicitly optimize for the benchmark itself, creating "peaky" performance rather than broad, generalizable intelligence.

Standard AI benchmarks are an engineering tool for measuring performance. A more scientific approach, borrowed from cognitive psychology, uses targeted experiments. By designing problems where specific patterns of success and failure are diagnostic, researchers can uncover the underlying mechanisms and principles of an AI system, yielding deeper insights than a simple score.

OpenAI's evals team is looking beyond current benchmarks that test self-contained, hour-long tasks. They are calling for new evaluations that measure performance on problems that would take top engineers weeks or months to solve, such as creating entire products end-to-end. This signals a major increase in the complexity and ambition expected from future AI benchmarks.

Silicon Valley now measures the intelligence of large language models like ChatGPT by their ability to play Pokémon. The game's complex mazes, puzzles, and strategic decisions provide a more robust and comprehensive benchmark for modern AI capabilities than traditional tests like chess, Jeopardy, or the Turing test.

Issues like 'saturation' and 'maxing' reveal a fundamental flaw: benchmarks test narrow, siloed abilities ('Task AGI'). They fail to measure an AI's capacity to combine skills to solve multi-step problems, which is the true bottleneck preventing real-world agentic performance and the next frontier of AI.

The gap between benchmark scores and real-world performance suggests labs achieve high scores by distilling superior models or training for specific evals. This makes benchmarks a poor proxy for genuine capability, a skepticism that should be applied to all new model releases.

Traditional AI benchmarks are seen as increasingly incremental and less interesting. The new frontier for evaluating a model's true capability lies in applied, complex tasks that mimic real-world interaction, such as building in Minecraft (MC Bench) or managing a simulated business (VendingBench), which are more revealing of raw intelligence.

When models achieve suspiciously high scores, it raises questions about benchmark integrity. Intentionally including impossible problems in benchmarks can serve as a flag to test an AI's ability to recognize unsolvable requests and refuse them, a crucial skill for real-world reliability and safety.

A flawed or unsolvable benchmark task can function as a 'canary' or 'honeypot'. If a model successfully completes it, it's a strong signal that the model has memorized the answer from contaminated training data, rather than reasoning its way to a solution.