Well, OK. They put in 61,000 lines, the entire first 12 seasons of the Simpsons's. Lines such as
{'previous': 'Marge Simpson: Ooo, careful, Homer.',
'character': 'Homer Simpson',
'line': "There's no time to be careful."}
and then used it to generate more Simpsons dialogue. With a data set so well matched to the goal, the algorithm barely matters. Remember that recent paper, "Copy is all you need."?[1] This is the ideal case for that approach.
I'm thinking more in terms of loading in the detailed product descriptions and maybe manuals from a catalog, then letting users ask questions about how to do things with the products.
