|Abstract:||This paper shows how underlying forms are learned by an OT on-line learning algorithm. The proposed algorithm is stupid: it processes one form at a time under one grammar at a time. Stupid is good and effective: no extra learning mechanisms are required than the ones already involved in a general grammar learning model of OT (e.g. Boersma 1997). Interpretation of incoming forms and constraint reranking as a reaction to error detection is enough. Surface and underlying forms are connected via faithfulness constraints, and underlying forms are connected to meaning via lexical constraints. It is shown that the learner can acquire grammar and underlying forms concurrently, and creates an economical lexicon.