The Sequence AI of the Week #891: Prompting a Spreadsheet : Inside Google’s TabFM for Tabular AI

Machine learning has a dirty secret: the models actually running enterprise operations aren't flashy transformers, they're gradient-boosted trees chewing through CSV files. Churn prediction, fraud detection, credit risk scoring — the stuff that keeps businesses running still relies on XGBoost pipelines, and the process behind them hasn't shifted much in ten years. Load the table. Engineer features. Cross-validate. Tune hyperparameters. Watch the AUC creep up. Then do it all over again for the next dataset.

Google Research just released something aimed squarely at that grind. TabFM is a foundation model built for tabular classification and regression, and it does something the old workflow never could: it makes predictions on tables it's never encountered before, in a single forward pass, without any training, tuning, or feature engineering. Feed it the entire problem at once — training rows, test rows, everything — as one massive prompt, and it produces an answer. It's in-context learning, except instead of text it's working on spreadsheets.

The lineage here matters. TabFM comes from the same team behind TimesFM, Google's time-series foundation model that ended up being deployed far more widely than most research releases ever are. That's not a coincidence — it's a playbook. Take a domain that's traditionally been dominated by bespoke, per-dataset modeling, and try to replace it with a single foundation model that generalizes across problems instead of being retrained for each one.

What makes this notable isn't just the technical trick of feeding tables through a model as a giant context window. It's the target. Tabular data is unglamorous compared to language or images, but it's also where an enormous share of applied machine learning actually lives. If TabFM's approach holds up outside a lab, it's chipping away at one of the most tedious and repetitive corners of the field — the part that never got much attention because it wasn't exciting, just necessary.