#char-rnn

The story of the chimney see Santa baby, and blood and joyous so world and joy and good will to see Santa baby bore sweet Jesus Christ Fa la la la la la la, la la la la la la la la.

King of toys and hippopotamuses [sic] full of the light of that stood at the dear Son of Santa Claus He was born in a wonderful christmas tree

Run, run Rudolph, run, run Rudolph, run, run Rudolph, run, run Rudolph, run, run Rudolph, run, run Rudolph, run, run Rudolph, run, run Rudolf the new born King

Some of them were actually pretty good:

Cherry kiwi

Peach belle

Cocoa & cream

Berry blossom

Melon mellow

Rose cherry

Some didn’t work as well:

Sugar snakle

Cherry coof

Stranger cake

Peanut butter beer

Coconut crapie

Some were really weird:

Coconut lick calangus

Armid pit

Butt bear bears

Sticed reaming rout

Peppermint catered picklemong

Spf 25 with greamint tit butter

Some were just inappropriate:

Pink poon

Strawberry shit hitt 

Coconut cream coot

Strawberry shite (It generated this one many times)

Banana butherfucker

For some reason, much like @lewisandquark​‘s neural network generated candy hearts, my neural net also seemed to generate a lot of bear-related flavors:

Strawberry bear

Peppermint toof bears

Coconut with bear

Coconut & smone bear

Butter bear

Bummy bears

Berry bear

Coconut bear (unflavored)

Spf 15 with bear

no what you need is one of these costumes

The Skypug

Hungry Boats

Mid wonka

Burderous bread cat

Holy Cheesarenda

Moth fairy

A magicial slice

Fall wearing monster

The Godfish

tiny neural net running on my laptop, trained on user-submitted halloween costumes

Back by popular demand, here are some AI-generated drawing prompts to use in this, the spooky month of October!

Midway through 2020, people started suggesting that I train a neural net on 2020 headlines, and I was skeptical that there would be enough weird ones to make a decent project. Then 2020 continued to be 2020. We started to get headlines such as:

Mysterious alien-like monolith discovered in Utah desert What you need to know about 'murder hornets' The Mystery of The Platypus Deepens With The Discovery of Its Biofluorescent Fur Famous Vienna hotel turns to drive-in cake

And by the time we got to December, with more bioluminescent marsupials, mysterious monoliths, and the Galactic Federation, my course was clear.

I decided to use GPT-3, whose internet training data was all collected before October 2019. In other words, before 2020 happened. I gave GPT3 a list of 2020 headlines and let it add more text to the list. Having no data from 2020, would it even treat these as real headlines?

A lot of the generated headlines were from the natural world, in part because my list of examples tended to favor those. I really can’t tell if these are attempts to do novel but realistic headlines, or to completely goof around.

Swarming bears are given deadly slingshots by Russian hunters Good news / Bad news about crows in Burlington schools When Killer Orchids Attack: How the Deadly Corpse Orchid Is Turning Up in U.S. Backyards Apparently There is No Good Photo of the "Melbourne Spider" Desert mystery - what is this rock that looks like a plane ticket? Fears rise of new dwarf hippo public relations disaster after rise in sightings Mysterious Origin of Monster Deep-sea Toads Solved What are 'dragon cats' and why they are getting hyped? Massive radioactive sinkhole continues to grow in Russia Why scientists believe the 'Killer Raccoons of the Pacific Northwest' are responsible for this kill Lycoperdons, the tiny deadly puffballs, are on the march again From deep in the Earth, darkness “boils” to the surface A sassy tardigrade previews new Doctor Who

With other generated headlines, it looks more like GPT-3 assessed the 2020 headlines and went, “Murder hornets? Yeah okay and hellhounds too.” (This is clearly an algorithmic error; hellhounds are MUCH more likely to be found at the Denver International Airport.)

Proof that a hellhound is living at Los Angeles Airport has been provided in the photos below First naked bogman has been found out walking the great British countryside Reports of a '10-foot tall penguin' roaming about on the Family Islands in the Solomon Islands are investigated Albino green sea monster filmed ... at the wrong time Scientists discover the alien ant farm under the Antarctic ice Lizardman: The Terrifying Tale of the Lizard Man of Scape Ore Swamp

I like its science/technology headlines. So many good little bots.

Researchers fear our solar system is being interfered with by alien space lasers Mars rover finds only artificial organic matter on surface Scientists Recreate Giant Liopleurodon In Lab 'Lost' NASA space shuttle repair robot comes back to life after four decades in complete darkness Hey, Curiosity Rover, Don't Go Chasing Shadow Bugs Small drone takes on six killer sharks in Shipwreck (Video) Mysterious Hole Observed On the Sun How did we miss this massive alien spaceship lurking on an abandoned asteroid? Scorpions on the dark side of the moon NASA: Has Earth 'Halted' And Stopped Spinning - If What NASA States is True, We Must All Celebrate! NOW!

i only showed the most interesting sorts of headlines; plenty were just ordinary news from 2019 or earlier. Others were new, technically, but not exactly news:

Breathing Is "Supercharging" Brain Voyager 1 images: See contents of interstellar space Scientists Take Photos In See-Through Scuba Gear, and This Is What They Got Scientists do science… … and then write about it! Scientists say climate change deniers should sit on a beach until it is underwater

Subscribers get bonus content: I tried another experiment in which I tried to produce more targeted headlines, around the theme of… tumbleweeds. (It turns out there were a LOT of dramatic tumbleweed headlines in 2020, due to their habit of descending on various places in apocalyptic numbers).

I’ve used various neural networks to generate recipes, to varying degrees of success. My earliest recipes were generated with char-rnn, which had to learn everything - spelling, punctuation, words - entirely from scratch. Its recipes were terrible (Swamp Peef and Cheese, anyone? Or Chocolate Chicken Chicken Cake?). Later when I started using GPT-2, I was using a neural net that had already seen a lot of recipes online during internet pre-training, and its recipes looked a lot more realistic (although not entirely realistic - there was that recipe for Black Forest Cake that suggested pouring the batter into a toaster, or the jello recipe that starts with “remove all internal rinds”).

GPT-3, an even more powerful text-generating neural net, has seen so many recipes during its internet pre-training that its recipes are boringly plausible. Bake its Black Forest Cake recipe and you’d probably actually get cake at the end, even if it’s not the best cake in the world. Its recipes aren’t funny at a glance, which is why I haven’t posted any of them - until internet linguist Gretchen McCulloch suggested that, rather than have GPT-3 generate its own recipes, I have it rate char-rnn’s recipes. Readers, it worked.

I prompted GPT-3 with two example recipes from allrecipes, each followed by 3 of their human-written reviews. Then, for the third recipe I inserted one of char-rnn’s recipes:

Chocolate Chocolate Cake

1 cup sugar 1/4 cup butter 1 cup margarine 1 cup cocoa 1 teaspoon vanilla 1 cup margarine 1/2 cup pecans, chopped 1 teaspoon vanilla 1 egg yolk 1/2 cup sugar 1/4 cup chopped pecans 1 cup sugar 1 teaspoon vanilla 1 cup cocoa 1 teaspoon vanilla

Combine flour, sugar, baking powder, salt, salt, and salt in a large bowl. Cook over medium heat, stirring constantly, until mixture boils and stiff peaks form. Add milk and cook until mixture is completely melted. Remove from heat and stir in sugar and sugar. Add eggs and vanilla. Beat until smooth. Spread in batter. Sprinkle with pecans.

Yield: 1 servings

And then I let GPT-3 add reviews. Would it mention the excessive lakes of vanilla? The complete lack of flour? The foolhardy reliance on the heroic efforts of a single egg yolk to hold things together?

oxtailsandturnovers Rating: 5 stars 03/04/2007 delicious chocolate recipe!!! and the perfect balance between cocoa and chocolate. and they sure were LONGNECKS! big time rave here. lazy w birthdays w friends and family, try this recipe for something chocolate sweet the ppl will talk about over and over.

AKEMIYUNA Rating: 5 stars appraisal 12/03/2008 Excellent for the kids Halloween party. Everyone got so thrilled by seeing something new on Chocolate in sponge. He even likes nuts separately but it will taste good..

BYOMFYABF Rating: 2 stars 10/01/2008 I could only give it a 2 as I had to make changes to it due to ingredients I had. First, I used Dove Dark Chocolate cocoa instead of Hershey's, no nuts, and the mashed bananas and caramel topping were a must. I did use coffee instead of the hot water, and dark margarine instead of regular, therefore I could not use light whipped cream, but I made homemade whipped cream. It had WAY too much flour in it, but i guess it could have been me measuring heavier than I should have. I also only used 2 cups of flour. it was my first bundt cake, so it was huge!!!

Here’s another batch of reviews, this time of a more chaotic char-rnn recipe:

Simple Soup Alaborate Soup Pie

6 each chopped white perserves; ** (optional but ring) 1/2 cup chopped fresh marjoram 1/2 teaspoon dried thyme 1/3 teaspoon salt 1/2 cup thinly sliced onion 1/4 cup onion (chopped) 1/2 teaspoon garlic; minced 2 small carrots 6 can dry white wine 1 tabasco sauce 1 tablespoon sugar 1/2 cup flour, whipped topping 2 tablespoon chili powder

Marinate: Serves 4. Whip and short under boiling. Add the sliced mushrooms and peppers and stir to coat. Stir the sugar and carrots. In large bowl, beat the egg to make a 1 1/2 tablespoons of the flour. Add shanks and stir it well. Spoon into a frying pan before serving.

Recipe By : DonajehLi Koubmanda

Yield: 8 servings

JEFFREY97389 Rating: 2 stars 03/03/2006 Even though everyone said this was great, I thought it was just 'OK!' The perserves gave a strange flavor to the sauce.

MATTJANDTRACER Rating: 2 stars 01/02/2009 This recipe sounded good. When it came out it tasted a bit like a casserole and not a stew. I used venison instead of deer and pork. I liked the way it turned out, but it wasnt very French.

Hammy Rating: 5 stars The g iving this as a gift as it takes about 10 less minutes but using American cheese and not blended with parsley flakes, lime-flavored gelatin, which is my husband's favorite dish from SuperValu. Just happens to be the creamy sauce. Don't worry about a pie crust.

In fact, the more chaotic the char-rnn recipe, the more chaotic the reviews tend to be, as if GPT-3 has picked up on the chaos and is trying to somehow produce reviews that could plausibly be written by whatever weird humans are visiting this mangled recipe website.

Geilded Peanuts & Crabmeat Basil

3 beef, boneless chops, or thyme, 1 lamber and flince of fresh diced in diagonal 1 parsley 2 large leeks (for garnish) 3/4 teaspoon sesame oil, seeded and peeled 1 any jeans'l-tortilla 1 pepper or cinnamon 1/4 cup baesly pinto beans. (Crushed and thawed; minced fine you recipe) 1 tomato base 1 fat in the last 5 inch plates 1 garlic clove; crushed 1/4 cup dijon mustard

In a large bowl set aside. Remove the sides from the heat and saute with a fork and cover and cook on Low for 6 to 6 hours.

Yield: 6 servings

Morales Rating: 0 stars 03/30/2008 I had high expectations for this recipe, but it was an outright flop. I don't know if the measurements were somehow incorrect or what, but I ended up with a watery mess that certainly wasn't crab salad. My girlfriend, who is a new chef liked it fine, but I really wish that the recipe didn't claim that it made six servings when it really made enough for one, if that's what we were going to call it.

Carole Rating: 5 stars Glorious sobs! This recipe will be delicious for a minumum of 1½ Hours to maximum 2½ hours Time to bring lots of fresh sized potatoes in the house or grow in the age may become quite famous

Amanda Rating: 3 stars 11/13/2009 My friend is having a birthday party and will celebrate with Camping. She's in charge of the dishes and wanted to serve food at the campfire. I'll 10th grade, my easy information for her party. It will not sound strange anyway test this task with a note to work rationally.What to take trees, portable tables for chairs, the main meal, food. I personally love the taste of it but has anyone seen a shepherd (wasn't any spices), oysters. It is sufficient

It does kind of make sense that GPT-3’s recipe reviews aren’t directly complaining about the utter weirdness of char-rnn’s recipes. Probably very few of the recipe reviews in GPT-3’s training data involved baffled complaints about nonexistent or repeated ingredients, or apocalyptic levels of recipe failure. So GPT-3 has learned that even the critical reviews tend to fall within a certain sentiment range. It’s interesting that it tries to emulate char-rnn’s primitive neural net glitchiness - I’ve noticed before that GPT-3 seems to be capable of pretending to be a much simpler neural net.

Subscribers get bonus content: After generating a few reviews, GPT-3 would sometimes go on to generate new recipes (such as “BBQ Cheeseburger Salad” which contains lime jello, whipped topping, and sliced black beans). It may have been deliberately trying to make them terrible to fit in with char-rnn’s recipes.

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.