AI & ML · Sep 20
Vals Raises $40M to Make AI Benchmarks Harder to Game
Andreessen Horowitz is backing Vals after the startup argued that older AI tests no longer measure modern model capability cleanly.
- Treat model benchmarks as evidence, not gospel, especially when public tests drive marketing claims.
- Choose evaluations that mirror your domain tasks, latency budget, and acceptable error patterns.
- Watch independent benchmarking tools as part of the AI deployment stack, not just PR infrastructure.