capital · May 21, 2026

Google DeepMind researcher Lun Wang resigns and criticizes AI evaluation methods

Lun Wang, a senior researcher at Google DeepMind, resigned and posted on X on May 19, 2026. In the post Wang thanked Google DeepMind for an amazing chapter and highlighted evaluation challenges for future AI models. Wang argued that current benchmarks excel at assessing existing models but fail for upcoming self-evolving models crossing new capability regimes, such as those involving emergent abilities or grokking. He called for investment in predictive, self-evolving evaluations as the upstream bottleneck affecting training, safety, and scaling. Wang linked to a blog post titled Your Evals Will Break and You Won't See It Coming.