What happened
A team at an unnamed company built a system that treats data selection and model architecture as connected problems rather than separate ones, showing that simply making models bigger produces worse returns unless you simultaneously improve what data goes into training. In practice, this means search results on e-commerce platforms improved measurably — a 1.7% jump in purchases and 2% in total sales on real users — by rethinking both what the system learns from and how it learns it.
Why it matters
This documents a shift in how industrial AI systems scale: the past decade treated model size as the main lever (bigger always meant better), but this work shows the bottleneck has moved to data quality and fit. If this pattern holds across search, ads, and recommendations — the three biggest commercial AI workloads — then companies optimizing only model scale are leaving money on the table and smaller competitors with better data pipelines could outperform them.