SAP RPT-1.6 is a tabular foundation model. It does not chat and it does not generate text. You give it a table in which some cells say [PREDICT], and it fills them in, using the other rows of the same table as examples. That is all it does.
Inspired by the LinkedIn post of Philipp Herzig showing RPT 1.6 playing Snake against Jev, I asked a playful question: can a model that only knows how to complete tables play Tetris? Not with a Tetris model trained for weeks, but with the stock model on SAP AI Core, the same model you would use to predict delivery delays or payment dates.
It can. Over two 100-piece games, RPT-1.6 cleared 38 and 35 lines. The hand-tuned heuristic it learned from cleared 38 and 39. No weights were changed. Everything RPT knows about Tetris arrives in the request, as 64 example rows per move.
This post explains how that works: how a game move becomes a table, what goes into the context, and why a table model can make good moves at all.
