In 1992, Andrew Blass published "A game semantics for linear logic", the first attempt to give meaning to proofs in linear logic in terms of winning strategies for a game. However, the composition of his construction was not associative, the two possible ways to compose three strategies were not necessary equal. Most of the subsequent works saw the non-associativity as an issue that needed to be corrected. In this talk, after explaining the Blass games in details, I will embrace their non-associativity, reveal their actual structure by using Munch-Maccagnoni's duploids, a non-associative model of effects, and show what it tells us about games.