The AI revolution is optimizing for the wrong things

Frontier AI models are improving at coding and agentic tasks while their performance on general prose is declining, as evidenced by measurable regressions on client-specific writing benchmarks after model upgrades. Maz Ahmadi urges enterprises to create custom evaluation suites tailored to their workflows instead of relying on public benchmarks, noting that only 37% of firms see EBIT impact despite 44% scaling AI enterprise-wide.