@ct_bergstrom@emilymbender I'm careful about arguments from incredulity like this. Even if I can't describe an algorithm for me to do this, I couldn't have described one for distinguishing cats from dogs using a list of pixels either. And there's stuff about few-shot and zero-shot translation using statistical regularities common between languages (https://arxiv.org/abs/1710.11041 and other papers by Artexte). Enough for me to not hurry to a conclusion even if it doesn't fully answer the challenge.