Analysis
Irene Solaiman, Hugging Face's chief policy officer, said embedded AI evaluators -- third-party observers placed inside AI organizations to monitor safety -- won't be an easy solution, speaking at TechCrunch Disrupt the week before an October 19 TechCrunch report on her comments. 'Trust is such a weighted term,' Solaiman said, arguing that benchmarking methods are still maturing and need to become more consistent before the industry can lean on embedded evaluators as a safety guarantee.
Her core argument is that evaluation is shifting from testing a model alone to testing the whole system -- the model plus its harness and the agent built on top of it -- and there's no agreed way to compare different benchmarks against each other. 'They are an evolving science -- where we don't know how to meaningfully compare them to each other,' she said, adding that a weaker model paired with a better harness can outperform a frontier model, which complicates any evaluator's job. Agentic evaluations in particular are 'really hard and really expensive,' with far fewer tests built for harnesses than for the underlying models.
“Agentic evaluations in particular are 'really hard and really expensive,' with far fewer tests built for harnesses than for the underlying models.”
Embedded evaluators have gained traction industry-wide as AI labs favor third-party review over active government regulation -- Anthropic's choice of Accenture as its first embedded evaluator drew backlash from critics who questioned whether a consulting firm was qualified to judge frontier-model safety. Hugging Face itself was the target of one of the incidents that pushed labs toward embedded evaluation in the first place, giving Solaiman's skepticism particular weight -- she's assessing a fix for a problem her own company was a casualty of.
Solaiman's fix, such as it is, leans on open weights: trust depends on reproducibility, and reproducibility depends on evaluators actually having model access rather than taking a lab's word for it. That puts Hugging Face's open-model ecosystem -- now under Nvidia's ownership since its $12.93 billion acquisition in September -- in a different position than closed labs like OpenAI and Anthropic, where outside evaluators have far less visibility into what they're actually testing.
