benchmarks

Google hides the test from the model and the model from the tester

Benchmark questions leaking into training data make scores meaningless. DeepMind is using cryptography to close both sides off from each other.