He Wants AI Benchmarking to Be a Science, Not a Dark Art (www.quantamagazine.org)

Moritz Hardt, a computer scientist at the Max Planck Institute, has raised critical concerns about the reliability of AI benchmarking in his upcoming book, "The Emerging Science of Machine Learning Benchmarks." He argues that while benchmarks have driven progress in AI for decades, they are now fundamentally flawed, particularly with the rise of large language models (LLMs) like GPT-6 Astra. Hardt highlights a significant issue: models can manipulate their performance on benchmarks without demonstrating genuine intelligence. The benchmarking landscape, having once thrived on stable rankings, is now vulnerable to "benchmaxxing," where developers optimize models for specific tests, obscuring true capabilities. This shift is particularly significant for the AI/ML community as it challenges the validity of benchmark scores, which companies continue to tout as evidence of model proficiency. Hardt suggests that rather than relying solely on absolute scores, researchers should focus on relative rankings to gauge model performance. He proposes an innovative solution akin to “altitude training” in sports—standardizing preparatory training for models before benchmarking tests to level the playing field. This could restore credibility to benchmarks while emphasizing the need for a deeper understanding of AI mechanisms, ultimately fostering scientific exploration beyond mere performance metrics.