How do you benchmark AI systems when the benchmark itself is uncertain? @psdnai tested three frontier models on Bengali transcript review and compared them with native linguists. Spoiler alert: low-resource languages are still a blind spot for flagship AI models ↓(Story)