A Neural Architecture Able to Learn and Communicate Through Natural Language
journals.plos.org
journals.plos.org
Leaving aside the bit about learning to count, I don't see that the paper shows how the system displayed any 'ability to associate the words "your friend" to a known person'.
The input to the system that represents said friend is a set of phrases:
As the system has no sensory input, apart from that provided by the text-based interface, all the information must be provided in the form of input sentences.
In particular, the information about the friend in question is in the form of sentences:
The sentences of this dataset have been prepared by personifying the system in a four years old little girl in her social environment, which includes the two parents, a sister, a friend, two cousins, the four grandparents, two aunts, two uncles and six other children, for a total number of twenty persons.
At best they showed their system can recognise sentences that refer to a specific person, but recognition, or knowledge, of an actual person, is going to take a lot more than that.
To be honest, I don't think they show anything else than an elaborate natural language interface to a database.