People can learn a new concept and use it compositionally, understanding how
to "blicket twice" after learning how to "blicket." In contrast, powerful
sequence-to-sequence (seq2seq) neural networks fail such tests of
compositionality, especially when composing new concepts together wit