Post by Vlad Larichev

Congrats, SAP! I can’t stop testing their new foundation model – and it’s a remarkable achievement, a real milestone for industrial AI 👏 The new SAP RPT-1 (Relational Pretrained Transformer) brings the foundation model concept to relational business data. Instead of text or images, it works directly with tables – the format that drives almost every enterprise and industrial process. Where GPT-style models process text sequences, RPT-1 processes tables – using 2D attention across rows and columns to understand relationships between variables like vibration, temperature, and maintenance strategy. Simply put: in a GPT-based setup, you’d have to serialize/translate your table into text string: 📑 “Equipment 1: vibration 2.15, temperature 71.94, ... RUL = 180.” That breaks structure and you lose the whole context of the tabular data! Try to read a table just as one endless line of text, and you will quickly see the problem! RPT-1 takes the structured table directly – numeric, categorical, and text columns remain structured. It learns in-context from a few labeled examples and predicts instantly, no retraining needed. It’s a foundation model that generalizes across thousands of downstream business tasks – 𝗳𝗿𝗼𝗺 𝘀𝗮𝗹𝗲𝘀, 𝘁𝗶𝗰𝗸𝗲𝘁𝗶𝗻𝗴 𝗿𝗲𝘀𝗼𝗹𝘂𝘁𝗶𝗼𝗻, 𝘀𝘂𝗽𝗽𝗹𝘆 𝗰𝗵𝗮𝗶𝗻 𝗳𝗼𝗿𝗲𝗰𝗮𝘀𝘁𝗶𝗻𝗴 𝘁𝗼 𝗮𝘀𝘀𝗲𝘁 𝗺𝗮𝗶𝗻𝘁𝗲𝗻𝗮𝗻𝗰𝗲 – all within the structure of enterprise data itself, predicting needed values from the context, with no training! If this direction scales, it could redefine how we bring AI into ERP and manufacturing ecosystems. Amazing work, Dr. Philipp Herzig, Boris Dmitriev, Jonas Budewig, Dominik Mark Metzger, Antoine Chabert, Bernhard Wörsdörfer at the SAP research team: Marco Spinac, Marek Polewczyk, Maximilian Schambach, Sam Thelin and many others! 👏