THE EVAL YOU CANNOT TRUST IS YOUR OWN

Summary from the TLDR IT newsletter

Enterprise AI evals fail when vendors can see POC test questions. Teams need a permanently held-out eval that vendors never see. Google DeepMind's double-blind frontier-model evaluation shows the model: neither side saw the other's evaluation assets. Cleaner lab benchmarks alone won't fix enterprise adoption failures.