AI AI Weirdness
Now that GPT-3 is capable of writing entire personality quizzes (sort of), could it write an entire blog post for me?
I prompted GPT-3 with “Welcome to AI Weirdness! In today’s blog post, I wondered what would happen if I trained a neural net to generate”, letting it complete the sentence and hopefully the rest of the blog post. Result: it has weird ideas about what would make an interesting post.
Welcome to AI Weirdness! In today’s blog post, I wondered what would happen if I trained a neural net to generate numbers in the range [-1,1].
As you might expect, the neural net started generating numbers that were more than a little odd: 0.9286401128496419 0.06491227721694329 0.9879906435756054 0.745677946535383 0.679450170988766 0.1065339560745793 -0.2977483969426606 0.05673549481890381 0.4304206725753945 0.7544217043730188 -0.8960721608636865 0.7224255864247317 -0.2139997613483644 0.0295771285597738 0.619233475195753 0.05578493537857566 -0.630712498931699 -0.0506476255464803 0.3256973044703828 0.0988207888549295
Welcome to AI Weirdness! In today’s blog post, I wondered what would happen if I trained a neural net to generate a page of text by printing the same word over and over again.
Welcome to AI Weirdness! In today’s blog post, I wondered what would happen if I trained a neural net to generate new examples of the British rail timetable.
The other completions it suggested tended to be even less interesting: long technical preludes about hidden layers of neurons and training data while never getting around to posting the results, or else posts in which the neural net was generating images that we never got to see.
I decided to get more specific, writing the topic and preamble myself so I could get straight to the neural net generated neural net generated results. (Yes, this is meta enough that I’m beginning to get confused). And, in a final twist that I almost wish I hadn’t tried, I wrote that the neural net’s results were weird. (First paragraph is my prompt, and the text that follows is verbatim from the neural net.)
Welcome to AI Weirdness! In today’s blog post, I trained a neural net to generate new kinds of cake by giving it thousands of existing recipe names. Unfortunately, the neural net was terrible at it, creating totally terrible-sounding recipe titles. Half of them seem to contain non-food ingredients. Here’s a list of some of the strangest titles:
Dump With Brown Ink Shot Show Ball
Dr. Huge Crust
Daily Beige Coarse Sugar
Cheese Yellow Cheese
Huge Aluminum Easy Nuts Beans Pancakes
Butter Cabbage Cake
Lemon Tough Compote Rice Cake
Minced Pepper Chocolates Muffins Sugar Pancakes
Orange Tart Vanilla Coffee
Spaghetti Jam Caramel Cake
Sticky Meat Orange Stew
Honey Honey Lemon Lemon Honey Lemon
I’m not sure how I feel about all this. Because I (and so many other people) have used neural nets to generate so much weird stuff, it’s now possible for me to use a neural net, trained on blog posts about the work of previous simpler neural nets, to pretend that I have trained a smaller neural net to generate weird things.
And it’s giving me weird things because I’m specifically asking for them. Here’s what I got when I gave it a preamble saying the neural net cake names could pass as human, and then again (at the same creativity setting) when the preamble said these were classic human cakes. As before, I supplied the first paragraph, and the neural net wrote the rest.
Welcome to AI Weirdness! In today’s blog post, I trained a neural net to generate new kinds of cake by giving it thousands of existing recipe names. It actually did a really good job! I wouldn’t be surprised to see most of these on real recipe blogs. Here are some of the most interesting:
Sunset Strawberry Cake
Seaweed Chocolate Cake
Spaghetti Caramel Cake
Butter Lemon Cake
French Honey Cake
Rainbow Strawberry Cake
Mozilla Grape Cake
Chocolate Heaven Cake
Chocolate Jelly Cake
Cherry Lemon Cake
Tobacco Lime Cake
Black Coffee Lemon Meringue
Welcome to Janelle’s cooking blog, in which I look at the most commonplace recipes in American cooking. Today, I thought I’d list some of the most frequently-baked cake recipes, the ones that anyone in the Midwest has probably had dozens of times. If you want to blend in, look no farther than these recipe titles:
One-Bowl Chocolate Cake
Lemon Cake Mix Cake
Golden Butter Cake
Angel Food Cake
White Velvet Cake
Moist Chocolate Cake
Hot Milk Chocolate Cake
Hot Milk Banana Nut Bread – I didn’t realize that this one fit the “extremely common” category, but it does, oh yes, it does.
Lemon Chiffon Squares – Makes 8 servings
German Chocolate Cake
Seven-½ Layer Cake – Need to make one now
Cheesecake – No sugar necessary – makes 8 servings
Although the neural net doesn’t completely succeed in creating realistic content, it’s clear that “neural net weirdness” is a writing style it knows how to produce on command. What does it mean when the neural nets are weird because we’ve asked them to be, not because they’re trying and failing to be ordinary? Maybe the creative work is now to figure out ways to nudge AIs into being weird and interesting rather than producing inane imitations of the most ordinary human writing.
After all, left to its own devices, GPT-3 would be writing A.I. Weirdness posts about numbers between -1 and +1. Neural nets have terrible taste.
Become a supporter of AI Weirdness to get bonus content: the favorite cookie recipes of famous Star Wars villains!
My book on AI, You Look Like a Thing and I Love You: How Artificial Intelligence Works and Why it’s Making the World a Weirder Place, is available wherever books are sold: Amazon - Barnes & Noble - Indiebound - Tattered Cover - Powell’s - Boulder Bookstore