Most AI breakthroughs involve text, images, or video. But huge portions of real decisions — credit scores, medical diagnoses, fraud detection — run on tabular data: rows and columns, like a spreadsheet. Each row is one observation; each column is a feature.

Here's the surprising part: deep learning doesn't dominate tabular data the way it dominates everything else. A technique called gradient-boosted decision trees — nothing like a neural network — consistently matches or beats deep learning on tabular benchmarks. Leading theories point to the fact that tabular datasets are often small, mix feature types, and contain no spatial or sequential structure for neural nets to exploit.

This is why getting foundation models to work as well on a spreadsheet as on a sentence remains genuinely unsolved — and why the business stakes are so high. Tabular data is what most enterprises actually have.