Don't do drugs.

looks like a cinnamon roll but could actually kill you: nakia

looks like a cinnamon roll and is actually a cinnamon roll: shuri

looks like they could kill you and can actually kill you: okoye

I’ve trained this open-source neural network framework on a variety of datasets, including recipes, Pokemon, knock-knock jokes, pick up lines, and D&D spells.

Now I give you: training a neural network on the complete list of Dr. Who episodes. (Big shout-out of thanks to blog readers Anya and Dan, who provided the dataset!)

The input dataset was simply the list of episodes, with no extra information given to the network - it has no idea even if this is English, or words, or a list rather than a paragraph or code.

Pretty much immediately, however, it figures out that “The” is important, as are the letter combinations “Do” and “Da”. The following is output from a very early generation, with the “temperature” (think “daringness”) variable turned to a very safe setting:

The Dont The Dant The Dant The The Dant The The Dont

If increase the temperature, forcing it to be more daring, it will try other letter combinations, but won’t give up on using “the”. Sometimes it’ll pile on a whole cascade of “the”s, in a fit of either panic or pique.

The Dinn The Sons The Ardeao The Areir The Wed on the The The Ans

Then, it finally learns to spell “Dalek” and never looks back.

Safe setting: (Daleks. Stick with the Daleks.)

The Sires of the Daleks The Argass of the Daleks The Arges of the Daleks The Wire of the Daleks The Argoss of the Daleks The Argass of the Daleks The Argas of the Daleks The Daleks of the Daleks

Adventurous setting: (Nonsense words and Daleks) The Wirs of the Arooss of the Daleks The Pas of the Amse The Aaression Ware The Amios of the Lale The Fist of the Daad The Girg Bans

In later generations, it has learned the tiny input set by heart, and on the safe setting is able to reproduce the actual episodes word-for-word, in order. On the adventurous temperature setting, I’m able to force the neural network to make spelling mistakes, and its attempts to move back toward some spot in the original list make for some interesting new episodes.

I give you: Dr. Who episodes produced by a rather annoyed neural network.