More benchmarks and more voices in AI testing

A post argues that a single benchmark measures only a slice of what an AI system can do. It calls for broader testing across environments, outputs and levels of autonomy—from simple prompts to full worlds, and single turns to continually improving agents. It also argues that subject-matter experts and everyday users, not just a small group of AI researchers, should help define what “good” looks like, supported by shared, open evaluation infrastructure.