Vals, a startup founded in 2024 and backed by Andreessen Horowitz, aims to revolutionize the way AI models are benchmarked. Recognizing that many existing benchmarking systems are outdated and easily gamed, Vals focuses on creating more reliable and industry-specific evaluations. After securing seed funding from 8VC and Bloomberg Beta, the company recently completed a $40 million Series A round led by Andreessen Horowitz, underscoring its rapid rise and growing influence in AI evaluation.

The company’s approach involves keeping test materials confidential to prevent AI developers from training models specifically to ace public benchmarks. Instead of general knowledge tests, Vals measures how well AI models perform complex, real-world tasks across various sectors such as law, finance, and coding. This method aims to assess not only a model's positive capabilities but also potential negative consequences if such AI were widely deployed, reflecting a deeper concern for real-world impact.

Rayan Krishnan, Vals’ 25-year-old co-founder and a former intern at Palantir and Microsoft, developed this model after observing that traditional academic benchmarks lag behind rapid industry advancements. Vals is expanding its coverage to include more specialized fields such as mental health, cybersecurity, biosecurity, and adherence to international laws like the Geneva Convention. Their unique evaluation process is likened to how students pay to take standardized tests, helping AI companies identify weaknesses and improve over time.

Vals is experiencing significant growth in both revenue and staff, with revenue increasing eightfold over the past year and its team tripling from eight to 25 employees. The company plans to expand further, opening a larger office and hiring more personnel. Vals has also launched programs offering evaluations to federal agencies and intends for its benchmarks to become essential tools for AI firms, especially as these companies prepare for public offerings and investor scrutiny in a market increasingly dependent on AI technology.