Gradient boosting
A classical machine learning method that still beats large language models at predicting values in spreadsheet-like data.
Gradient boosting builds a prediction by stacking hundreds of very simple decision rules, each one trained to fix the mistakes of the ones before it. It has been around since the late 1990s, it trains in seconds on a normal laptop, and for tables of rows and columns it is still the method to beat. Popular implementations include XGBoost and LightGBM.
This matters because it is the clearest example of a job where the newest AI is the wrong tool. Research published in 2026 confirmed that frontier language models lose badly at tabular prediction, and that they fail in ways no classical method explains. If your question is “predict this number from this spreadsheet”, reach for the old tool.