Who Grades the AI Models? | Ben Horowitz & Rayan Krishnan

As AI models continue to improve at a rapid pace, the methods used to measure their capabilities are quickly becoming obsolete. This creates a significant challenge for companies investing heavily in AI, as they struggle to accurately assess which models perform best for their specific needs and whether their investment is truly yielding value. This episode features Reyan Krishnan, founder and CEO of Val.ai, in conversation with a16z’s Ben Horowitz and Jennifer Li, to explore the complexities of AI evaluation, the limitations of current benchmarks, and the critical role of independent third-party testing.

The Problem with Public Benchmarks

Reyan Krishnan recounts the founding of Val.ai in early 2024, driven by the realization that existing public benchmarks were insufficient for measuring model progress. He highlights a fundamental issue: the tight relationship between building new generation systems and developing new evaluation mechanisms. Often, improving one necessitates improving the other. Krishnan explains that while many interesting models were coming to market, it was becoming increasingly difficult to ascertain what was truly new and capable. This led to the conclusion that a third-party company was needed, solely dedicated to building high-quality evaluations and benchmarks to discern the true capabilities of these new models.

Ben Horowitz raises a pertinent question: why can't the labs building these models do this themselves? They possess the deepest understanding of their models and what capabilities are missing. Krishnan acknowledges that labs do build many great benchmarks, which drive model progress. However, he points out a critical issue: the self-reporting nature of model capabilities. He cites the example of Meta's Llama 4, which underperformed on Val.ai's private benchmarks but showed incredible capabilities on major public benchmarks. This disconnect between self-reported performance on open benchmarks and actual performance on higher-signal, private benchmarks underscores the need for independent evaluation.