I don't trust benchmarks. So I built an alternative (unbenchmark.com)

A new platform called Unbenchmark has emerged, designed as an alternative to traditional benchmarks by focusing on user reviews of AI models. This initiative highlights a growing demand for more nuanced evaluations of AI systems, as it allows users to share their experiences and insights. Unbenchmark aims to enhance clarity in communication regarding model performance, particularly with updates like Fable 5.1, which appears to exhibit both improvements and limitations, such as making unsolicited changes. The significance of Unbenchmark lies in its departure from score-based evaluations, emphasizing user feedback for real-world applications of AI models. Users report varied experiences, with some praising the model's efficiency in specific tasks while criticizing its delegation capabilities and verbosity. This nuanced approach to model assessments could reshape how the AI/ML community evaluates performance, pushing for more dynamic and context-driven understandings rather than relying solely on standardized benchmarks. Overall, Unbenchmark reflects an evolving landscape where user-centric evaluations hold the potential to drive improvements in AI tools.