In this article, he shows with concrete examples how Google Translate falls short, and then offers two criticisms:
* Feeding more data to current models won't bring them any closer to understanding, since understanding involves having ideas (including ideas about the state of the world), and this lack of ideas is the root of all the problems for machine translation today. His examples are powerful evidence of this limitation of current state-of-the-art machine translation systems.
* Current machine translation systems make no attempt to go beyond the surface level of words and phrases. These systems merely discover statistical regularities that relate words to other words at multiple, hierarchical levels of composition. In Hofstadter's words, "there's no attempt to create internal structures that could be thought of as ideas, images, memories, or experiences. Such mental etherea are still far too elusive to deal with computationally." He is right.
That said, AI researchers are aware of these limitations, and are exploring possible ways to overcome them. An early, crude example of such research is the multi-modal model of "One Model to Learn Them All" (https://arxiv.org/abs/1706.05137), which is trained to learn to perform multiple image-recognition, language-translation, image-captioning, speech-recognition, and language-parsing tasks at the same time, using representations that are shared by all tasks.
While these early research efforts fall far short of the kind of "understanding of the world" Hofstadter shows is necessary for human-level language translation, it's encouraging to see AI researchers actively looking for ways to move beyond the mere discovery of 'hierarchical statistical regularities' that relate words to other words.
This is an exciting time for AI research.