How far has AI come in a year? In 2018 I generated cocktails using textgenrnn and char-rnn, two neural networks that learn to imitate text. I trained them on 1000 cocktails that Beth Skwarecki collected for her neural net cocktails bot, and watched the neural nets struggle to make sense of what they were seeing. They came up with nonexistent ingredients and had such limited memory that they didn’t know how many times they’d already used a particular ingredient. One cocktail, the Black Banana, was a behemoth containing a staggering amount of creme de cacao.

black banana
 * 1 1/2 ounces gin
 * 1/2 ounce dry vermouth
 * 1 ounce cream
 * 1 ounce creme de cacao
 * 1/2 ounce creme de cacao
 * 1/2 ounce rum
 * 1/2 ounce lemon juice
 * 1/2 ounce creme de cacao
 * 1/2 ounce triple sec
 * 1/2 ounce grenadine
 * 1/2 ounce creme de cacao
 * 1/2 ounce amaretto

 - fill cocktail shaker with ice.
 - add gin, vermouth, and pineapple juice.
 - shake.
 - strain into a collins glass.
 - add coffee liqueur, creme de cacao, and cream.
 - shake.
 - strain into a chilled cocktail glass.
 - garnish with a lemon slice.

But in 2019, the neural net GPT-2 was released by OpenAI, who trained it on 40GB of text from the internet. GPT-2 runs on the big processors on Google Cloud (through Max Woolf’s gpt-2-simple colab notebook), and has much better memory than textgenrnn or char-rnn. And when I trained it on the cocktails, I found that GPT-2 could keep track of its ingredients and figure out that it was supposed to just use them once.

But its cocktails don’t seem to be any good. And in some cases, they’re very very bad.

=== BIG WHITE FIZZ
* 1 ounce Grand Marnier
* 1 ounce gin
* ½ ounce brandy
* 1 ½ ounces cream
* 1 teaspoon sugar
* 1 teaspoon Tabasco sauce
* Dash MACARONA

- Fill cocktail shaker with ice.
- Add Grand Marnier, gin, and cream.
- Shake.
- Strain into a collins glass.
- Fill with crushed martini glass.
- Garnish with pepper jack cherry.

(NOTE: do NOT make this cocktail)
(Further NOTE: apparently MACARONA is a popular Egyptian pasta dish, which is an INTERESTING choice for a cocktail)

The neural net apparent obsession with creme de cacao also seems to be alive and well. As is the inadvisable handling of glassware.

AI’S FAVOURITE

1/2 ounce creme de Moly
1/2 ounce triple sec
1/2 ounce melon cordial
1/4 ounce creme de cacao
1/4 ounce creme de cacao

- Heat creme de cacao and triple sec until very hot.
- Add melon cordial.
- Stir.
- Strain into puddled whipped cream glass.
- Fill with creme de cacao.

The neural net can also tend to go overboard with the garnishes.

DEATHOH SHAPIRO

* 1 ounce ginger ale
- Add ginger ale, sugar, and salt
- Stir.
- Pour into a chilled martini glass.
* * 2/3 cups crushed ice
- Fill cocktail shaker with ice.
* Add crushed ice.
* Fill with crushed ice.
* Add crushed ice.
* Pour into a collins glass.
* Garnish with a maraschino liqueur.
* Serve with a cherry.
* Garnish with a strawberry.
* Enjoy this high-fidelity saucer with a cherry.
* * Garnish with a maraschino liqueur.
* Garnish with a maraschino flower.
* Garnish with a maraschino nutmeg.
* Garnish with a mint sprig or an orange slice.
* Garnish with a sweet cherry.
* Serve with a nutmeg slice.

And GPT-2 almost has TOO MUCH memory for cocktails, in that it assumes that cocktails must be related to the cocktails that came before, with the result that when it has generated a weird cocktail by accident, the rest get even weirder.

AI’S FAVOURITE CAFE - MIXED MIXED MIXED MIXED FLAVOUR - MIXED ALUCE MIXED MIXED MIXED FLAVOUR - MIXED MIXED MIXED MIXED MIXED MIXED MIXED MIXED FLAVOUR - MIXED MIXED MIXED MIXED MIXED MIXED MIXED MIXED MIXED MIXED MIXED MIXED MIXED MIXED MIXED MIXED MIXED MIXED MIXED MIXED MIXED MIXED MIXED MIXED MIXED MIXED MIXED MIXED MIXED
- ¾ ounce tequila
- Float a dash of grenadine
- Add tequila, grenadine, and grenadine
- Blend until smooth
- Serve chilled in a microwave-safe glass.

=== DINOSAUSTE - TUBE IN SCREW
- Fill the microwave with ice.
- Add this ice to the microwave.
- Add an orange slice from one of the cans to the mix.
- Stir.

=== THE BEAR
- Dust and gold with powdered sugar, and serve with fruit slice.
- Garnish with a large flintlock.

=== ROYAL SUFFICIENT
- Fill a goblet with ice.
- Add ice.
- Squeeze to thicken.

I experimented a bit with ways to prompt the neural net into making ever stranger cocktails, and one of my favorites was to give it the sentence “Captain Picard leapt from his chair, exclaiming loudly. The ship had come out of nowhere.“ The neural net still remembers how to write fanfiction based on its general internet training, but now it has a tendency to twist the story around toward cocktails. Old-fashioneds, for some reason.

Captain Picard leapt from his chair, exclaiming loudly. The ship had come out of nowhere.

He cut through an old, stone-encrusted old-fashioned glass spigot and filled it with wine. Showing no passion, he shot a flintlock into the glass.

Impish, the old-fashioned glass!

Reading a bunch of these is almost like reading some form of experimental poetry. Become an AI Weirdness supporter to get them as bonus content! Or become a free subscriber to get new AI Weirdness posts in your inbox.

Subscribe now

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

Subscribe now