On a whim, I decided to see what would happen if I let my Wikipedia-title neural network be a bit more creative, by increasing the "temperature" setting to 1.25. This basically means that, at each character in the title, it's more likely than it normally would be to choose a less probable option. (For comparison, the output that I've posted here in the past is typically run with a temperature of 0.75.)
Sometimes that means the model has a tendency to form longer words than average:
Usidro-McNamandridigatico
Genia oxivor transpooneepurregits
Croxidus mestachaidpresswards
Jefferschavalwaivian de Nube
Puertopula Bigocomusepislopiskemacy
Guaddecylanzacesvake (festerument literatic)
Cathosadewren's bobrosmastionatable
Romantomeals of temporated bosetownshipfeed Strixpbellistoneer-pool
Progresmeradopouç cabinalist gubpwil-talementadwritalenous
Sometimes it means putting letters together in combinations that arenât quite so common:
Wighffsoner: Hudgerlania (2)
Dinistpwillapruc (cizidian)
Tshorvquocnes Computer F.C.
Nipreijasvi (Fzez Tazilical)
As the last entry from each of those sets suggests, sometimes it just throws a diacritical mark or other foreign letter in there because it can:
Richware (Neee Nake): The BaĂtoran
Sometimes, it gets a little bit too eager with those diacritics (and with the letter "z", for that matter):
ĆizOĆĄĆ„Ärz, Dasyf 20 Rep. 240
And as always, it comes up with lots of intrinsically funny new words:
Plade McOntonisation, Greece album
NGC video dynastergy relations
Roback of Hap Brinkles (disambiguation)
Borber stape (Beautime & United York)
On Sword Alfo (Paris, Indoxidation)
Marchy Sparking, Thomaceptions
105th Kentucky Corpignation
Amailo, Wrogdam Conseboopbool
