Last week I trained a neural net on 1000 candles, and soon it was producing scents like Frozen Styrofoam, Volcanoes Comfort, Lemon Lime Decay, and Friendly Wetsuit. We have to just imagine what these would smell like (or in some cases, try not to imagine). But what if the neural net could describe them?
Amy Pollien offered to help out, developing a web scraping tool that could browse candle sites and collect the candle names and descriptions. After just 72 descriptions, though, she had to admit defeat. “The tool can’t tell the difference between the product description and hordes of enthused human users gushing about how “the scent of seagulls takes them back to fantasies of fresh wash hanging on the line” and I guess I’m OK with how I couldn’t anticipate that.”
I trained GPT-2 on this ridiculously tiny list of candle descriptions, but let it learn for only a few seconds before halting - if I let it go longer, it might memorize the examples. I didn’t expect much of a neural net trained under these circumstances, but when I asked it to generate text, I got things that - well, they WERE candle descriptions.
An Oyster Roast
An amber sea creature mumbles and takes its sweet perfume. The smell turns heads and then descends into the sea. The sea is rich and salty and warm and delicious.
Rowan's Flora Garden
Floor to root (entry the candle is shelled), raisins, citrus, earnut and strawberry inside lush, bouquet-lined rows.
Despite its tiny bit of training data, the neural net was able to draw on boatloads of similar text that it had seen during its previous general online training. “Oh, you mean THIS sort of stuff.”
I’m not exactly sure where it got some of this stuff from.
Freshly Marbled Water, Only 2% Sulfur. Perfect for when your face emits ultraviolet light, this decadent beret and floral-laden concoction adds a warm undertone of fresh scent, with notes of water, lemon essence, wabi wax, and bergamot.
Sometimes it seemed to be doing perfumes. Or health supplements. Or beer. Or mixed drinks.
Traditionally associated with bad post-apocalyptic stories, this drink conjures up images of a dark lab, smouldering and broken. A smoky, fruity scent radiates from the lily of the valley, and with every whiff of citrus and tropical fruit on hand, each smell quenches the 1000 notes of chrysanthemum.
The neural net wasn’t supposed to be trained on French, but saw just enough French that it thought it could fake it. It is bad at French.
Beuve Seguin comme vertu sous filosie Ât un cherche platage de mer. Salud leur firmaiente grise grise. Saucé ou bien gran doth fait.
Its favorite scent is clearly patchouli. I saw so many candles like this.
Woodsy in Rose
The ghostly scent of sage overtones at sunset. A rich and earthy lily of the valley. Captured in amber and petal with notes of jasmine, patchouli, patchouli patchouli, and leather patchouli, this ancient Roman incense-laden incense candle measures 3.5 by 2.5 feet (with a base of chrysanthemums), and lends a patchouli-tinged touch to any room smell when opened.
These are so characteristic of neural net text - it’s good at matching surface features like vocab and turns of phrase, but experiments like this show how it has no deeper understanding of what it’s doing. It can come up with new ideas, but it doesn’t know the difference between an interesting idea and a terrible one.
Warm Blood Water Swatch.
This evening's cold wash brings warmth and pleasantly sweet notes of fresh lime, vanilla, and warm warm water across a warm and damp back.
AI Weirdness supporters get bonus content: more neural net candles, including a “candle” that’s more like a recipe. A recipe that tells you to add ½ a pebble and sacrifice 1-2 lb. short straight-cut pecans. Or become a free subscriber to get new AI Weirdness posts in your inbox!