Quick note: For over two years now, we’ve been running The Sequence without sponsors. Literally every week we receive tens of inquires to sponsor The Sequence. As a passion project that I have been running for over five years now, we are trying to keep The Sequence true to its ethos of technical depth, objectivity and unique content to help you navigate the world of AI. The best way to help us on that journey is to subscribe below. We understand is not for everyone but appreciate the support.
There’s a running joke in machine learning that the field’s most valuable model isn’t a transformer at all — it’s gradient-boosted trees fit on a CSV. For all the noise about frontier models, the workhorses of enterprise ML are still XGBoost pipelines predicting churn, fraud, and credit risk on rows and columns. And the workflow around them has barely changed in a decade: load the table, engineer features, cross-validate, tune hyperparameters, repeat until the AUC stops moving. Every new dataset means starting the ritual over.
Google Research just took direct aim at that ritual. TabFM, released a few days ago, is a foundation model for tabular classification and regression that produces predictions on tables it has never seen, in a single forward pass, with no training, no tuning, and no feature engineering. You hand it the whole problem — training rows, test rows, all of it — as one giant prompt, and it answers. It’s in-context learning, but for spreadsheets. And if the name sounds familiar, it should: this is the same team’s playbook from TimesFM, the time-series foundation model that quietly became one of the most-deployed research artifacts Google has shipped. Understanding TabFM really requires understanding that lineage, so let’s start there.
The TimesFM prelude
Read more