Reporting by XinZhiyuan
Editor: Yuan Xie La Yan
【New Zhiyuan Guide】Do you think that the love words on various cards and candies on Valentine's Day are the same? Every year, I go to the personal blog of Ai internet celebrity Janelle Shane to collect her strange love words generated by AI routinely: "Love two thousand wild boars", "Hit my clothes", "Ants can stay".
Valentine's Day is here.
Whether it is a 10,000-year-old single house that is always lonely, a lover waiting for the other half to offer a routine annual salute, or an angry wolf who hates the materialization of human nature in commercial festivals, you will expect to see an overwhelming number of mass-produced single flirtations today.
The love story of Valentine's Day is boring and almost universal. So, if you want the ghost step dance under the moon to take the usual path and have a little strange love story, how to do it?
There is a big sister who can write codes: take AI to run.
In 2018, it began to use personal computers to train AI to talk
Freelance researcher Janelle Shane, the main job is a laser scientist working in an optical equipment company, and her side job is to engage in her hobby of neural network AI training, write AI popular science books, and be an INTERNET celebrity in the AI industry.
In 2017, when GPT was not available, she took her MacBook and trained a simple neural network AI to write Harry Potter fans, create new character names for the Star Wars Universe and Dungeons & Dragons games, and generate new Pokémon cards.
At the end of 2017 - Valentine's Day in 2018, she began her famous habit of using AI to generate a batch of Valentine's Day heart-shaped candy/card love words the following year.
Around Valentine's Day, these products, known as "candy hearts", will be sold in Europe and the United States for a while. They're tiny, with some brief Valentine's Day-related messages written on them. Heart-shaped candies usually have room for only a few characters, so write something like "I love you" or "Call me~ or "I'm yours!" Stuff like that.
In order to create the "Love Talk Bot" AI, Shane initially collected 366 love stories about the valentine's day heart-shaped candy that are actually on the market. Feed these raw datasets into a neural network, let the AI identify data patterns, and then use those patterns to generate new mock love stories.
Well, the result is indeed quite novel. However, it can be seen that AI is far better than real people in the technology of showing love and seducing people. The resulting love story products are also far from the commercial quality that can be filled with candy and sold for money on cards.
Although it can't be sold for money, the result is really very strange.
Some of the AI-generated love words have reached the sweetness standard:
"LOVE BUN"
"You are babe" (YOU ARE BABE)
DEAR ME
"Cute Kiss"
"My Bear" (MY BEAR)
Some of them almost mean:
"YOU ARE IT"
"Heart ME" (HEART ME)
"FANCY MY HERO"
More went in strange directions:
"ALL HOVER"
TEAM BEAR
PIN A FACE
"Bog LOVE"
"I Honker" (I HONKER)
Others have gone into truly bizarre territory:
"Love 2000 HOGS YEA"
BEAT ME TAME
"Stank LOVE"
"Sweaty Poop" (SWEAT POO)
"Sweat PEAR"
"CHERT FACE".
Some have also entered the field of adult puns:
"SWEET POLE"
"MEAT MATE"
"You A GOO" (YOU A GOO)
Lick
"LOOK BIG"
"My Little Slut" (MY HAG)...
According to Janelle Shane's response in an interview, she found it very interesting to use AI to generate hooks to people's words, because the algorithm did not understand the specific meaning of words, and did not know why changing the meaning of a letter would change. Simple AI at the time would only learn patterns for data arrangement and arrange letters according to those patterns.
However, with AI as a mirror, it can be seen how funny human beings can be when they court, after all, the original data words are all excerpted from commercially available real heart-shaped candies.
In 2020, use GPT-2 to generate gibberish love words
The 2017 Neural Network AI, which generated Valentine's Day love stories, had zero experience training in English and could only learn the original 366 words — and it still didn't know which letters to avoid in certain combinations.
At the end of 2019-Valentine's Day 2020, Janelle Shane experimented with GPT-2 to generate love words.
At the time, GPT-2 was arguably the strongest AI on the market in natural language processing, with over 1.5 billion parameters and 35,000 copies of training texts crawled, and it excelled in various tasks of "predicting the content after a given text".
Although GPT-2 hadn't learned anything about Valentine's Day cards at the time (though it might have seen a list of cards online), Shane used Talk to Transformer to add existing heart-shaped candy and valentine's day card content data to see what it would output.
Shane knew, though, that GPT-2 wasn't a neural network dedicated to generating heart-shaped candy love words. What she did was a bit like walking up to someone and shouting," Hot guy! Cool man! Sweet couple! Call me! Magical Boy! Forget it!" A hodgepodge of these words, any real person who hears them will feel confused.
But in fact, this is similar to shouting "lubricant" to the neural network AI and expecting feedback from the RESULTS of the AI operation. To be honest, most natural language processing experiments are so nonsensical.
So does this neural network AI know what it's doing? Should not know. After GPT-2 outputs some full-text capitalized text, it continues to display other types of text. From these other texts, it is clear what it wants to output:
lyrics
Band name
Interesting little knowledge of animals
Campus massacre records
Ringtone tune
Even if GPT-2 is given a clue and marked clearly in the hint that this is some Valentine's Day-related information, GPT-2 still doesn't seem to know what it's outputting. This type of text may be rare in its training data.
Another clue is that the heart-shaped candy love stories generated by GPT-2 are usually long and nonsensical —it has no concept of length limits. For example, the following:
Insertion is difficult and there is a real need for low-density sturgeon
God bless the undead team
Hot stuff, my body is
Exude overflowing love
A single face sheet on the floor of the dance floor
The message from the crypt said it was very happy to see you send me a friend message
I wonder what number it is today
Be wary of our bottom layer
How to dress like a bat
American ocean cabbage delight
Chocolate banana cheese chunks
In that case, can I invite you to eat cookies?
The resulting long sentences are terrible, but GPT-2 is quite successful in generating love sentences. For example, there are still sweet ones:
HEARTED TREAT
LOVING HORN
Dancing on A LOAF
The weird ones are:
ANTS CAN STAY
DOOMED