What happened
Researchers show that a standard metric for measuring how efficiently AI vision models run on hardware (MACs, a count of mathematical operations) doesn't actually predict real-world execution speed, especially on edge devices like phones and embedded systems. They built a new vision model called LowFormer that optimizes for actual speed rather than operation count, and it runs faster across multiple devices while maintaining accuracy.
Why it matters
For years, AI engineers have optimized models using a metric that doesn't match reality — like designing cars by counting cylinder firings instead of measuring fuel consumption. This matters because it shapes which models get built, deployed, and funded; a shift toward metrics that match actual hardware behavior could redirect billions in AI infrastructure spending toward systems that actually run efficiently in production.