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

For over two years, The Sequence has operated without sponsors, maintaining its focus on technical depth and objectivity in AI. Google Research has developed TabFM, a foundation model for tabular data that enables predictions on unseen tables without training, tuning, or feature engineering, functioning through in-context learning similar to its predecessor, TimesFM.

• The Sequence is reader-supported and has operated without sponsors for over two years.
• Enterprise machine learning often relies on gradient-boosted trees for tasks like churn and fraud prediction.
• The traditional workflow for tabular data involves feature engineering, cross-validation, and hyperparameter tuning.
• Google Research’s TabFM is a foundation model for tabular classification and regression.
• TabFM makes predictions on unseen tables in a single forward pass, without requiring training, tuning, or feature engineering.
• The model operates using in-context learning, where the entire problem is given as a prompt.
• TabFM follows the approach of Google’s previous model, TimesFM, which was a foundation model for time-series data.
  https://bender.layer3.press/articles/b8e85f39-3a09-445f-8f85-b34b1fb9b54d