Hugging Face's Solaiman: AI evaluations 'an evolving science' we can't compare

The debate over benchmarking has grown as AI labs stake their reputations on evaluation results, but the methods behind those evaluations remain complicated and still maturing, per Solaiman. Reproducibility has long been one of the stickier aspects of benchmarking, and she noted the added question of how one test compares to other evaluations.

Embedded evaluators have been popularized over the last several months as labs grapple with a rash of safety incidents, and Hugging Face was famously the target of one of those incidents. Labs have pushed for the third-party solution instead of active government regulation, yet doubts remain over how neutral or effective those evaluators will actually be; Anthropic's decision to hire Accenture as its first evaluator drew backlash from critics who mocked the consulting firm's qualifications.

Evaluations are also shifting from the model alone to the entire system around it, including the harness and agent. Solaiman said agentic evals are really hard and really expensive, and that there are fewer evaluations for harnesses than for models even as better ones for the holistic agent are developed. She described the current evaluation space as a bit of a free-for-all, suggesting a more uniform industry-wide approach may be necessary if benchmarking is to remain a useful tool for leaders.