Benchmarks Don't Build Great Products. Engineers Do.

The AI industry faces a benchmark problem, not due to an excess of benchmarks, but because too many companies view them as mere accolades rather than practical tools. The primary function of a benchmark is to enhance a product, not simply garner marketing attention. A suitable benchmark should challenge engineers before it catches the eye of marketing.

If a benchmark fails to improve the product, it may not be fulfilling its core purpose. The AI sector has witnessed numerous benchmark announcements in recent months, with each release claiming state-of-the-art achievements. This surge of benchmark-driven releases often leads to cynicism, as companies unintentionally blur the line between measurement and mission.

At Backboard, they do not rely solely on benchmarks, favoring transparent measurements that are better than subjective claims. Benchmarks function as feedback loops, revealing strengths, weaknesses, and the impact of implemented changes. They deter self-deception about a product's progress and foster a shared understanding among customers.

However, benchmarks can be manipulated through several common methods, such as optimizing specifically for a benchmark, leaking evaluation data into training, cherry-picking configurations, and publishing only favorable results. These practices serve marketing interests rather than product improvement. The critical distinction lies in whether a company builds for benchmarks or benchmarks what it has built.

The former revolves around optimizing for a test score, while the latter prioritizes solving real problems first and then validating improvements through independent evaluations. Transparency in benchmarking is crucial, encompassing publishing methodologies, open-sourcing evaluation frameworks, and sharing logs and results. This approach fosters reproducibility, challenge, and continuous improvement.

Benchmarks cannot encapsulate every facet of an AI system, but they provide a standardized framework for comparison. Benefits include reaching a wider audience and sparking conversations with potential stakeholders. However, the focus remains on the sequence: first benchmark, then build, ensuring engineering practices align with real-world value creation.