via
https://ift.tt/35CJX1hlewisandquark:
People wonder about what it would be like if a super-intelligent AI decided to place all of humanity in a realistic simulation. I wonder what it would be like if the simulation were built by today’s AI instead - whose computing power is somewhere around the level of an earthworm’s.
Specifically, what would the pies be like?
In 2017 and 2018 I gave increasingly powerful neural networks about 2000 names of existing pie recipes, and their imitations included Caribou Meringue, Mur’s Tartless Tart, and Cromberry Yaas.
But there are now even more advanced neural nets, including GPT-2, many times larger and pretrained on millions of pages from the internet. When I used gpt-2-simple to train the smallest version of GPT-2 (145M) on the list of pies, even it was able to come up with plausible (and highly repetitive) pies, many copied from the list I gave it. I can’t blame it for being a smartass and giving me the pies right back - technically this is a perfect solution to its task of “name some pies”.
I have ways of messing with the neural net, however. One thing I can do is increase the chaos level of the text it’s generating (in this case, called the “temperature”), forcing it to go with its less-likely guesses. Another thing I can do is give it weird text to complete. When I’m asking it to generate more text, I can give it a few example pies and ask it to add to the list (the original list of pies I showed it was alphabetical, so if I give it apple pies it tends to respond with more apple pies). OR, I can give it much weirder text. If I give it a bit of Harry Potter fanfiction, it draws on its previous training and can continue the story. But, similarly to the version of GPT-2 I trained on crochet patterns, it tends to twist the stories around to its special interest. In this case, pie.
Prompt: Harry glared at Snape.
Harry glared at Snape.
“No, I’m not even talking about the boy here. You’re talking about the duplicate.”
“Fine. Then I’ll get the rest of the ingredients.”
So then, they got canned fruit and whipped cream and whipped cream and whipped cream and whipped cream and whipped cream and whipped cream and whipped cream.
But one thing was for sure: there was no trace of rum.
There was.
Prompt: Voldemort ran his fingers gently over Darth Vader’s shoulder.
Voldemort ran his fingers gently over Darth Vader’s shoulder. “C-Custard”
“C-Custard”
Once it has steered the conversation, the neural net transitions to listing pies. Although some of them retain a certain flavor.
Voldemort Crumble Pie
Mudblood Mincemeat Pies
Snape Pie With Cheese
And its other pies are indeed more creative, definitely not found in the original training set. They don’t sound exactly delicious to me, but maybe this advanced artificial intelligence knows something we don’t. After all, it read most of the internet, including the cooking blogs.
Hot & Sour Apple Pie
Caramelized-Zucchini Nog
Nachos From Maple Syrup Pie
Syrupy Chicken Pie
Camel Crunch Crunch
Mustard Raisin Pie
Peach-pickle pie
Pie-Bro From Pizza Particles
Iceberg Chiffon Pie, Famous
Finger-To-Finger Meringue Pie
Spiced Coconut Pizza Chiffon Pie
On the other hand, maybe just maybe the neural net doesn’t know what it’s doing. These seem like uniformly a bad idea.
Fog-Crust Pecan Pie
Tweezers, Olives, and Seafood Apple Pie
Frozen Custard Pie with Three Glorious Chilies
Half-Ton Pie Crust
Fog-Crust Pecan Pie
Pumpkin Ear Pie
Onion Cassette Pie
Tweezers, Olives, and Seafood Apple Pie
Frog-Fudge Pie
Ice-Egg Pie
Bedtime Oil Tart
Half-Tooth Nail Pie
Frozen Custard Pie with Three Glorious Chilies
Must-Fail Pecan Almonds with Pecan Almond Crust
Beef and Cheese Fried Ronald McDonald’s Walnut Chocolate Pie
These neural net inventions miiight be okay if we knew what they actually were. At the very least, approach with caution.
Applewool Pie
Brandywine Schnoches
Pumpkin Pie Dough - Oh Noe
Cowdy-Doo Pie
Dried & Squirred Pumpkin Pie
Morgue Pie
Temptation To Fill A Beef Pies
Woollewinkle Pie
“Vinegar Pie”
Meat Pie**
Stimulant Cherry Pie
Deep Dish Curb And Spoon Pie
Standard Apples-and-Moolah
My Meat Crust (Might As Well)
And these particular examples seem to have outed themselves as the work of a less-than-whollly competent AI.
Vessel With Raisin Bran
It’s Crunchy Famous
Pie Crust (Spinach Friend)
Food And Flavors Whoopie Pie
Valet Restaurant’s Chompin Park Pie
Medieval Pastries With Liquid Crystal
Tomato/Banana Pie As I Layered
Mr. Robot’s Sweet Potato Pie
Fluffy Cheese Pie #2 By BestMakesCherryToEggs
Ahh So Good ‘n Easy Meat Pie
Scheduled Coconut Cream Pie
Standard Pie (Panties & Puff Pastry)
Smiled and Cheerful Thanksgiving Pie, Famous
Tarte Barre “Oral” Pie
Robocoin Pie
For some examples of how the neural net managed its transitions from fanfiction to pies more or less gracefully (too long to fit in this blog post), enter your email here.
My book on AI is out, and, you can now get it any of these several ways! Amazon - Barnes & Noble - Indiebound - Tattered Cover - Powell’s
![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}