Janelle Shane is a research scientist who plays with neural networks like, textgenrnn. This one in particular, which she wrote about in the blog AI Weirdness, “tries its best to imitate any kind of text you give it.” Whether you give it a dataset of paint colours, band names, or guinea pig names, the network will try to predict appropriate combinations of letters and words that fit into a specific category. When she gave textgenrnn 1,228 American cookie recipes to predict dessert names, she received an interesting collection of outputs.
Based on these hundreds of existing cookie names, the program output titles like Low Fuzzy Feats, Walps, Fluffin Coffee Drops, and Grandma’s Spritches. Don’t those just sound delicious?
To the neural net, the names don’t mean anything. As Shane described, “it’s just picking letter combinations that seem likely to it.” Earlier when she fed textgenrnn a list of paint colours, the program suggested Stanky Bean, Stargoon, and Turdly…I’m not sure anyone will be looking to paint their bedroom Turdly anytime soon, regardless of how appropriately organized those letters are for a paint colour.
Though it’s fun to see what textgenrnn outputs for cookies, maybe we should stick to naming them ourselves for now, otherwise we may end up with some Hand Buttersacks.