
September 29, 2026 · 11 min read
AI Benchmarks: What the Leaderboard Hides
Founders reach for AI benchmarks at one moment: when they have to pick a model and want a number that makes the decision for them. It will not. Vals AI co-founder Rayan Krishnan says Llama 4 underperformed on his team's held-out private benchmarks while looking excellent on the public ones, where the questions and rubrics are open source. Benchmarks also saturate, go stale, and have shrunk to small samples with long rubrics. Here is how to read a leaderboard, and the only test that picks your model.
Read more →



