#textgenrnn

I’ve used neural networks to name all kinds of things - halloween costumes, craft beers, cats, and even guinea pigs. The weirder the starting set of names, the more a neural network’s creations might blend in (although cats named “Jexley Pickle” and “Big Wiggy Bool” might at least raise eyebrows).

How about the names of racehorses? The Jockey Club recently released a list of over 42,000 registered racehorses “in an effort to assist owners in identifying an appropriate name for their Thoroughbreds.” They don’t say whether it’s okay to use it to identify inappropriate names for Thoroughbreds instead. 

So I trained two neural networks, char-rnn and textgenrnn, on this tantalizing dataset. Some of the names they generated might pass for the names of real horses (which after all include horses named Three Little Birds, Spider Magic, Fierce Pajamas, and even a prizewinning horse called Cloud Computing).

She's a Babe North Storm Magic Creek Desert Cat Mary's Blade Fabulous Charm Frisky Joe Velvet One Dawn Forever Lord Power

(I decided to illustrate a few using BigGAN (via ganbreeder.app), another neural net that generates pictures. Unfortunately, “horse” is not a category of image it can do, so I decided to go with “horse cart”. There was an attempt.)

The fact that these names are not taken yet highlights the potential of neural nets for this important task. The next set of names, however, highlights their disadvantages.

Snow Motion Farthand Harren's Broke Bodyyhole Pond Wind Orcha Shuffleston Rapple Musty Pick's Lilver Ginky's Rental Suckee Habon Culf Q's Wart Aaging C Bon Warry L Z' W.

(For these illustrations I experimented with mixing in birds for more colorful horses - hornbill, and flamingo)

If you want your horse to be memorable, though, neural net names may be just the thing.

Party Can Devil Tina Exclusive Bear La Flumber Unbridled Dave Secret Sand Killer the Butterfly Pickle J Gator I'm Repent Evilish Kelly Snuckles Will Surprise Zegafish Pete Pat's Quick Hat Mr Fact Cowgitter Stark Moretowiththebotterfron

Left: “horse cart” plus a bit of “anenome fish”. Right: “horse cart” plus a bit of “great grey owl”. “Dog” crept in there because BigGAN saw a LOT of dogs during training.

This post from Janelle Shane was the most hilarious use of a neural network I’ve seen. As soon as I saw it, I started thinking of other datasets to feed it. Why not every burger of the day from Bob’s Burgers?

So I installed textgenrnn and started laughing my ass off at what it was coming up with. Here are some of the best. Enjoy!

I am really really excited to announce a thing I’ve been working on for more than a year now: there’s going to be an AI Weirdness book!!!!!

You can preorder it at these links:

Amazon - Barnes & Noble - Indiebound - Powell’s - Tattered Cover

When I was coming up with a title, I decided to see if I could get a neural network to help me out. Since they learn to generate text by following examples, my first task was to collect a bunch of book titles for the neural net to learn from. Of course, there haven’t been that many cartoon-filled AI humor books published, so I had to resort to other kinds of books.

Starting with a version of textgenrnn that I trained on Welcome to Night Vale transcripts, I then trained it on the full text of my book - that would give it a starting vocabulary that was an uncanny version of my book. But that version of the algorithm was generating paragraphs, not snappy titles, so I trained it for just a bit on another dataset I had handy - ice cream flavors. Here are the titles it generated.

Strange Serve Butter Chapters NOT Chocolate Top Chocolate Cream Chocolate People Cake

Although delicious, the titles weren’t quite as computer-focused as I would like. I took that same confused Night-Vale-AI-ice-cream neural network and trained on a list of computer algorithms.

Quickletely in a recognizy Strawbergendon's algorithm Kevin But algorithm Dang also

They tended to be either very wordy or very obscure - I’m not sure what I was expecting from algorithm names as titles. So I decided to take my cue from science fiction titles. I filtered a big set of titles from http://www.isfdb.org/ for length (so I wouldn’t end up with unwieldy suggestions thanks to titles like “The Philosophical Voyager in a Land Unknown to the Inhabitants of the Earth” and “Captain Megaloman and His Trusty Sidekick Squidley Save the World Once Again”, and then trained the neural net on these.

The Dead Death The Blood Bandland The Scatting The Tamelight of Winter Tablike

None of these titles had anything to do with humor, or artificial intelligence, or any of the things the book was actually about. So I started over with the ice cream neural net (I had to start somewhere, and I was hungry) and this time I filtered the scifi titles for length AND for a bunch of keywords like machine, robot, artificial, etc. Finally I was getting a bit closer.

One group of titles seemed to be about cute helpful robot friends, and that was nice. My book IS full of cartoons of helpful robot friends (other than the accidental murderbots).

Robot posting! No, Strange Friend Artificial Friend Little Strange Machine Go Strange Robot A Friend you?

But my book also talks about times when AI is NOT that helpful, times when it makes mistakes or solves the wrong problem, sometimes with serious consequences. I did get some generated titles that struck more of a scary note.

The Doom of Humans Get the Human Strange Murder Machine The Book of the Skin's Home! Hello, The Hole of Night The God and Oh, It's Humans The Million Murders Book The Blood Robot

These titles are needlessly scary, though. Another point of my book is that AI isn’t smart enough to come up with goals of world domination - it’s got the approximate brainpower of a worm. There were some titles that definitely brought out that aspect of the neural net.

The Book of Thinks The Human Found and Four in Room! The Artificial Lives, I really on the Jokes The Very Say Chairs Everything Moon! Still We Book of the Dave Robot Bookbook of Dog Machine

But another aspect of AI is that it can be delightfully, unpredictably weird. It can bring in all sorts of new connections that humans would never have thought of. My book’s a celebration of that weirdness. There were some titles that were indeed weird.

Horses of Pants Spoon: Vampire of the Book Handway Wormsels! Hello, Go Hover My Face of Monsters The Fighty Earthquake The Pig-Book of the Stars The World Human Shovel! By My Bunned Rocks Get Melted Zom! The Book of Pies High McDangs!

And of course there were the titles that were basically repulsive. Those were fitting, too. Today’s AI doesn’t have a sense of appropriateness or morality - it will accidentally do Very Unfortunate things because it has no clue there’s a problem.

Hello, You Colons History of Sweat The Sign of Fart! Strange Human Spoor! The Hole Attack! Hello, the poopers Humans Swearing Machine

So, did I end up using a neural net-generated title for my book? Yes, I did! You Look Like a Thing and I Love You is definitely the fault of a neural net. But it wasn’t generated by the neural net I trained on book titles after all. Instead, longtime blog readers may recognize this title from a different neural net: one trained on pickup lines. 

I hope you all will have as much fun reading my book as I did writing it!  There will be hilarious experiments, real-world examples, and illuminating cartoons. And yes, there will be plenty of giraffes.

You can preorder it here (preorders are super super helpful for letting stores and my publisher know that people are excited about the book, which helps it do well. If you’re planning to order it anyways, a preorder would be an awesome way to do it).

Preorder You Look Like a Thing and I Love You! Amazon - Barnes & Noble - Indiebound - Powell’s - Tattered Cover

So I’ve used neural networks to generate recipes in the past. They’re computer programs that can learn to imitate the data we give them, copying the way that humans drive cars, label images, or translate languages.

That is, they try to learn. They’re called “neural” because they have virtual neurons that work a little like the real neurons in our brains. Their virtual brains, however, are really tiny. Where a human has about 86 billion neurons, the neural networks we use today have hundreds to low thousands - think nematode worm or, optimistically, jellyfish. So they often struggle, and recipes, my friends, are one of those times. (you haven’t lived until you’ve seen a neural net’s attempt at “small sandwiches”)

I turned to cocktails, inspired by Beth Skwarecki’s cocktails bot (and helped immensely by the dataset she sent me).

The neural net’s first attempt was... an attempt.

I let it train for a little longer, and things got a little more recognizable, as it learned to kinda spell more ingredients.

It had, however, not learned when to quit. 64-ounce (2 liter) cocktails are not unheard of at this stage.

It continued to learn on its own, all without input from me, and eventually ended up with cocktails that were almost plausible.

Some of my other training attempts, however, did not go so well. At first the cocktail titles in my training set were in all caps, which confused the heck out of the neural net because capital letters were so rare that it didn’t see enough examples of what to do with them. It struggled with the titles.

You may also notice that the ill-fated all-caps attempt had a lot of repetition - that’s because I was using textgenrnn, which has a somewhat limited memory. It could learn to spell ingredients, but had no idea whether it had already added sugar and cream. When I switched to char-rnn for the lowercase recipes above, I could give it a memory of 50 characters, enough to cut down on repetition. The textgenrnn version, however, became strangely obsessed with creme de cacao. The less said about its cocktails, the better.

For more neural net cocktails (including custom-generated cocktails for any name you care to provide), check out Beth’s cocktails bot!

A while ago I made a bunch of new pies. Well, I didn’t *make* them because they were neural network invented titles and although it tried to imitate the list of pies I gave it, the neural net’s imitations are imperfect.

The neural network, after all, is a computer program with about as many neurons as an earthworm. It doesn’t understand what the ingredients are, or why some combinations don’t work. Some of its titles were intriguing, though. They sounded mysterious. Potentially delicious and/or magical?

Or maybe it just helps that they’re vague. I decided I wanted more like these. To help it along, I spiced up the pie dataset with the names of cookies and apple varieties from the 1905 edition of Apples of New York. I filtered the names for those that had possessives: Mcaffee's Nonesuch, Cornell's Savewell, Wile Ox's Winter (all apples), combined with Goldy's Dungeon Bars, Esther's Bracelets, and Fido's Rewards (all cookies). Then, to give it added old-school flavor, I added all the Dungeons and Dragons spells that had possessives as well (for example, Ivy's Irresistible Scent, Freedom's Toast, and Leomund's Tiny Hut).

I arranged the training data so the pies would be last (so they would be freshest in the neural net’s virtual mind). Then I gave it one single look at the data.

It turns out that I didn’t manage to prevent the neural net from coming up with bad ideas. Perhaps what I should have done instead was remove all the meat pies from the training data.

But some of the pies were exactly what I’d hoped for.

And some even went a little past “ancient” and into “legendary”

So there’s these computer programs called artificial neural networks that are good at imitating things. By seeing examples of what humans did, they can learn to translate languages, predict product sales, and even categorize text and images as innocuous or explicit (it has a lot of trouble with this last task, as it turns out).

One neural network I use, called textgenrnn, tries its best to imitate any kind of text you give it. I’ve given them paint colors, band names, and even guinea pig names and in each case their results are somewhat... mixed. (Paint colors called Stanky Bean, Stargoon, and Turdly, for example) The problem is that it doesn’t know what any of these words mean - it’s just picking letter combinations that seem likely to it.

This is what happened when I gave it all the cookies from a list of American recipes. This is what human cookies sound like to a neural network.

Now if you’ll excuse me, I’m going to go whip up a batch of Fluffin Coffee Drops.

Support AI Weirdness and get bonus content: More cookies, including the neural net’s strange obsession with “balls".

I did one of these for 2017 last year, so I thought it would be fun to recap some of my favorite aiweirdness.com experiments from 2018. Sure, financial forecasting, facial recognition, and delivering ads may be the most common uses for machine learning, but that doesn’t mean they’re the best uses.

FOOD

To me, this never gets old. Give a neural net a list of recipes and ask it to imitate them, and the resulting creations become a sort of surrealist performance art.

1. Cocktails

Key characteristics: a very short memory and an inexplicable fondness for creme de cacao.

2. Ice cream

The neural network got the cadence of ice cream flavor names, but never quite got that there were some words it should avoid. (Last year it had a similar issue with generating pies) Favorites that the Kealing Middle School Coders generated, and favorites that I generated with the training data they collected. 

3. Cookies

Generating cookies was an even easier task than cocktail recipes and ice cream flavors because the neural network didn’t even have to learn to generate real words. People ended up baking their own interpretations of many of these, and they were much better than anything the neural net could have come up with on its own. Human/AI partnership -- AI still needs humans to actually make the cookies!

KNITTING

4. SkyKnit

A project called SkyKnit had a neural network attempt to reproduce knitting patterns. Human knitters translated the resulting very strange instructions into actual knitted objects. The fewer of SkyKnit’s mistakes they fixed during debugging, the weirder the knitting became.

[Make Caows and Shapcho - invented by SkyKnit, knitted by MeganAnn]

ROMANCE

5. Candy Hearts

It turns out that neural nets can also imitate candy heart names, although giving these to your sweetheart may or may not have the intended effect.

6. Burlesque

Speaking of romance, I had a go at naming a new burlesque show with a neural net as well. The resulting post, “AI Does Not Understand Sexy” was temporarily (and ironically) censored by Tumblr as containing adult material. The first neural net-named burlesque show, “My Rear’s On The Sexy” played to a sold-out house.

My 2017 Farty Burlesque Adventure Sex Your Eye Out! Gourdraiser! The Stripper Stripper Dave Burlesque Show The Pants of Fame Burlesque Adventure

JOKES

7. Telling jokes

Joke-telling is one of those few areas where science fiction has it right about AI - neural networks do indeed have a tough time with humor. I trained a neural net on about 43,000 jokes and the resulting joke-shaped but not joke-containing things reminded many people of the jokes that five-year-olds tell.

How many engineers does it take to change a light bulb? A star an alligator and because they are bees.

What do you get when you cross a pirate and a little butter? A bathroom.

Why do seagulls fly over the sea? Because they know …………..

8. April Fool’s jokes

A neural net’s pranks are the kind that people will never expect, and will never understand.

Put marbles in the refrigerator. Place a pair of pants and shoes in your ice dispenser. Putting googly eyes on someone’s computer mouse so that it won’t work. Put marbles in the hand soap dispenser.

NAMING THINGS

9. Snake species

Just discovered a new species of snake and not sure what to call it? This neural net has you covered.

10. Robotics teams

When the neural net gets tired of wowing us with its culinary masterpieces and instead turns toward world domination, at least the name of its elite strike force will be interesting, judging by the names it suggested for these high school robotics teams.

One of my favorite things to do with machine learning algorithms is to get them to name things. Give a neural network (a type of machine learning algorithm) enough examples, and it will do an often decent (sometimes howlingly bad) imitation of their letter combinations and patterns, all without the slightest clue what any of it means. I’ve used neural networks to name craft beers, cats, paint colors, pubs, and even guinea pigs.

Here, I trained two neural networks, char-rnn and textgenrnn, on a set of 24,548 boat names, courtesy of Bradley Peterson and a sailor from San Francisco.

Behold the name of your next boat (or next self-aware starship?)

First, the decent/plausible ones:

Sky-Shark Sun Princess Windsprite Sandy Bear Catchwind Flowing Wind Windshark Wave Bard Sea Rainbow Windshare Second Sun The Windmower Sea Bug Miss Salty Bully Of The Water

And for the somewhat more eccentric sea captain:

Bear Storm Slow Shark Hugs Sea Hog First Ramen Sun Of Bear Little Boot Ship The Trick Wart Day Of The Grover Running A Goose Shadow Again Ken Dream Princess Cracking Dusting Hare String Shark Jelly Gall Fumper Huggy Tail Bear Goose Midnight Rat Pea Fury Song Fish Of The Dog Quiet Leg Cookie Sea Back Shark Salty Lick Wolfing Sun Nature Charging Lover

I'm kind of fond of the ones that really do sound like someone trained an AI to write them:

Nice Boat What Day Saily Sea Blue Of The Wind Merry Biggles The Fun Big Ditty Sun In The Sun Stinker Delusions The Salt, How Are Sons Otto Smells The Bug Wolfs ! Trivial Pockets Wands Off! It.... Holeygum Arch & Navel Spocked

These are available at a discounted rate:

Stinky Day Silent Wind Snot Runner The Shanky Wind Bad Surfer Lick Musty Fin Princesside Maolly Fart Putty Mom Just Ship Aquamock Stinkbarr Drifting Furdy Some Dog The Liver Silent Rock Spanking Winds Sand Plow