Interesting. You could add counting and other jobs as well, eg ”how many times does value 8364 appear, and in which dictionaries?” A lot of the needle trained

Interesting. You could add counting and other jobs as well, eg ”how many times does value 8364 appear, and in which dictionaries?” A lot of the needle trained models clearly can’t reason with stuff everywhere in a long context even if they can retrieve. I would like to see some extensions that require some reasoning. I guess you could use words as dict vals and ask it what sentence a set of keys spells, and to answer the sentence with a set of keys. Lots of interesting possibilities.