Janelle Shane (@JanelleCShane) is the reigning queen of making funny, silly, and downright bizarre texts with neural networks. She's an electrical engineer who works with holographic laser beams by day, and plays with neural networks in her spare time.

Her projects use char-rnn, a type of neural network which she explains by linking to Andrej Karpathy's "The Unreasonable Effectiveness of Recurrent Neural Networks," which details how recurrent neural networks process the data fed into them, such as the complete works of William Shakespeare, and produce something that looks exactly like Shakespeare, but is a genuine fake.

Andrej Karpathy is currently the Director of AI at Tesla, and his blog and github page have been cited by innumerable data scientists, hobbyists, and others interested in learning about how neural networks work and how to set them up themselves for fun and profit.

Janelle Shane's projects started going viral this year with a popular one being a neural network that could design and name paint colors. Here are some of the best results:

Her conclusions: the neural network really likes brown, beige, and grey, and has "really really bad ideas" for paint names.

But if you think Clardic Fug, Burble Simp, and Stanky Bean are as weird as it's going to get, please read on.

What's in a name?

Names are short and pretty easy for a char-rnn network to re-arrange and re-combine into something that sounds vaguely plausible to us, so many of her projects have involved naming conventions.

Cheese Breeze and beer please

Training the network on 1,500 names from the My Little Pony Friendship is Magic Wiki, she had one experiment naming new My Little Ponies. With names like "Rainbow Dash" and "Fluttershy" already extant in the Ponyverse, this one was marked as a partial success. While plausible ponies like Sunshine Star and Glowberry were produced, there were also ponies like Cheese Breeze, Apple Ronch, and Groan.

Another similar project created craft beer names, with a delightful degree of realism. Look out for the Dang River IPA, Frog Trail Amber Ale, and the Sir Coffee Stout on your next trip to the neural network taproom.

May the farce be with you

Another project creating Star Wars character names unsurprisingly produced a lot of Siths at the lowest creativity levels, including "Darth Darth." And wasn't Darth Teen the villain in The Force Awakens?

She also created one for Star Wars planets (with extant names like "Tatooine," and "Hoth," anything is possible) and used @i_find_planets to flesh out their descriptions. If you haven't had your personal planet found, we recommend tweeting "Planet, please!" at @i_find_planets for a planet of your very own.

Foppin and Popchop

Shane's project to produce cat names was trained on several hundred names from a cat rescue in Alabama, and several thousand cats registered in Toronto, which should have created names following North American cat naming conventions.

But she first trained the network on the wrong data set, using a list of fantasy names by J. R. R. Tolkien, George R. R. Martin, and others, producing such exquisite cat names as Mankith, Belfine Bracken, and Grim Wyyne.

Eventually (when trained on the correct data set) this project resulted in such suitable names as Snox Boops, Foppin, and Mr Gruffles. Other cat names that she deemed less successful, but we must beg to differ: Sofa, Pope, and Pissy.

She also named guinea pigs for the Portland Guinea Pig Rescue. Meet Popchop and Fuzzable:

Popchop & Fuzzable via lewis and quark

If The Adventures of Popchop and Fuzzable isn't a buddy comedy in the making, what is?

The best of the rest

In addition to naming new sports teams, cars, and Pokémon, she created neural networks that could name 80s action figures and new musical acts.

The 80s action figures project created Thunderator, Mannosaurus, and Action Bun, which calls to mind a cute little bunny rabbit armed with a bullet bandolier, as well as Ninja Rat, the Mr. Splinter spinoff show which definitely had a TV spot between the Teenage Mutant Turtles and the Samurai Pizza Cats:

Most people think anything could be a band name, and with bands like Shpongle, Spoon, and !!! out there, it's hard to dispute these neural network names' plausibility: The Freights (which probably sounds like The Shins, but recorded from inside a boxcar rolling down a lonely track at midnight), Nighty Daggers (maybe something like the Arctic Monkeys? but with more stabbing), and Skins of Space (which was definitely a rejected name for the glam rock band that would become The Darkness).

She was also able to name metal bands with a huge amount of data (100,000 bands including genre and country of origin) and I personally can't wait to go see Death from the Trend, the Black Metal outfit from Croatia, next time they tour with the Russian Melodic Death Metal band Inhuman Sand.

Stranger Than Fiction

"Yer a wizard, Harry!"

The first experiment in neural network-generated Harry Potter fan fiction used a data set of 10,000 examples from Archive of our Own (or AO3), which produced very weird but still plausible stories such as:


"The Perfect Cow by alafaye
Severus and Hermione start a horcruxes"

"Birds of a Saturday by SasuNarufan13
Harry Potter is drunk and discovers he is an alternate universe."


Then to refine her slash machine, she retrained the network to work on the word level, and not the character level. Word-level allows for more grammatical sentences, longer memory, and generally more readable output, and is limited to using words that appear more than 200 times. Here are a few things that appear way more than 200 times: Happy Birthdays, Christmas, Harry Potter characters together in bars, and the phrase "more than they bargained for."

When the network is instructed "to play it Really Really Safe, and choose the most likely next word in each sentence," the output is, well, Harry Potter in a nutshell:


"A Hero’s Tale by orphan _ account | Harry Potter is a wizard, and he is a wizard. He is a wizard. He is a wizard, a wizard, a wizard, and a son. He is also a Slytherin, and he is a wizard. He is a wizard, and he is a wizard. He is also a wizard, and he has not been the one to be a father."

"A Hero’s Tale by 1001Angel | Harry’s life is turned upside down when he finds out that he is a wizard, and is a wizard."


The show must go on

A neural network trained to create Broadway productions, including closing and opening dates, produced plays that were not limited in their performances by the constraints of linear time, or the normal rules of decorum in naming conventions.

Results included a comedy entitled Butt, which ran for over 7 years and was only performed once, and a much more successful play called Fart, with a 4-year run and performance count of 23 times.

Wise or otherwise

Three more of her experiments relied on tricky human patterns of speech: proverbs, fortune cookies, and knock-knock jokes.

Many of the ancient proverbs sound like they were dreamed up by the Inspirobot and could pass: "No wise man ever wishes to be sick." While others revealed a strange obsession with oxen.

The fortune cookies produced almost no usable answers. But the knock-knock joke generator produced this laugh-out-loud gem that you will definitely want to use at parties:


Knock Knock
Who’s There?
Ireland
Ireland who?
Ireland you money, butt.

Mastchar-rnn chef

Saving the best for last, Shane's most hilarious project to date was trained on 30,000 cookbook recipes, and based on Tom Brewe's project to create recipes using a neural network.

The results sound like something out of the surrealist cookbook by Salvador Dali Les Diners de Gala, or from the Manifesto of Futurist Cooking.

Combine chunks and sprout clams

The network created disturbingly vague ingredients:

1 cup mixture
1 teaspoon juice
1 chunks

and oddly specific ones "that you could plausibly ask for at Whole Foods and act all disappointed when they don’t have any" such as "milked salt."

Which combined with dubious cooking instructions such as "Fold water. Roll into small cubes." and "Sprout clams; add vanilla." to produce the most improbable meals since bread in a can.

Just for fun, she once gave her cooking network the complete works of H. P. Lovecraft, and asked it to complete sentences, or start them, producing such spooky instructions as:

"Coat apple slices with strange things."

"Cook over medium heat until thickened and bubbly. Spoon over bizarre eyes."

"Sometimes, in the throes of a nightmare when unseen powers whirl one over the roofs of strange dead cities toward the grinning chasm of Nis, it is a relief and even a delight to make the soup."

The cake is a lie

Here's one recipe for a "cake":


BAKED OTHER LIE 1993 CAKE
appetizers, fish
8 rounds; chicken
¼ lb butter (soaked)
1 can tomato sauce (½ lb)
1 salmon steaks sauteed
½ teaspoon red pepper, chunked
1 tablespoon margarine or oil
Meanwhile, transfer the chicken breast to a serving platter and simmer for about 5 minutes, then lemon juice that has been stirring well; if on the side, as becomes warmed, carefully frost them with a sauce. Spread them and garnish with water or parsley.


The intrepid Jono Ellis actually baked a vaguely-chocolate-chip-cookie-related recipe created by Shane's neural network, with the secret ingredient of horseradish.

They did not follow the instructions, only the ingredient list, and said it made a very fine cake-like mixture:

"The horseradish is a subtle background flavour and the overall spicy, peanut-y, chocolate-y flavour is ace."

In an interview with NY Mag, Janelle Shane said she tried the recipe herself and

"It was the most horrible chocolate thing I have ever tasted in my life. I opened the oven and my eyes just watered. It was so bad."

She then reveals that she brought cupcakes of it to two different parties, and none of the guests shared Ellis' opinion on the palatability of the baked good:

"The two different parties that I took it to, I found out that somebody had quietly taken a bite out of one of these cupcakes, and abandoned it somewhere."

Super Deluxe also recreated one of her recipes, which no one was brave or misguided enough to try eating:

Taste test

In interview with the Daily Dot, Shane was asked whether she thought computers in the future could create recipes that would actually be good to eat:

“I could imagine a consciousness appreciating food even with no way of ingesting it—as long as they had sensors to pick out nuance and complexities the same way we might appreciate a symphony or a painting.”

Relying purely on the text of cookbooks, rearranged however which way the neural network pleases, has none of the nuance or complexity of taste bud sensors, but produces an absurdity that delights the mind more than a tasty snack.

And for creating or recreating things like humor or wisdom, the char-rnn algorithms do about as good a job as can be expected of a non-thinking entity; it's probably a good thing that computers haven't developed a sense of humor yet.

But maybe we don't need nuance and sensitivity for everything. For the things that really matter (like bands, cats, and craft beers), the proof is in the pudding.

Permalink