This tale is very closely based on real facts.
I have been working recently on a new project based on artificial intelligence and neural networks. Nothing too fancy: just a proof of concept to check how advanced neural networks really are, especially those involved with natural language and human knowledge.
The goal was simple: to set up a neural network able to express best wishes. She (from the very first moment our neural network was a she in our minds) should be able to greet us every day with a sentence.
We began by investigating what were the technologies available. We found one of them particularly promising: it’s a kind of neural network called LSTM, an acronym for Long Short-Term Memory.
In our case, as we were teaching her how to greet, we needed greetings, lots of greetings.
That was our first problem: where to find so many greetings. How many? At least 25,000. 25,000 sentences are a lot of sentences.
We walked the usual venues: we asked our friends and families to write some greetings for us, we ripped a few websites dedicated to inspirational quotes, we bought a lot of greeting cards… After a couple of weeks of really hard work, we looked at our corpus with joy and pride.
But when we counted our greetings we felt a little discouraged: 1,000. We had exhausted our ideas. How would we manage to get 25 times that? We had encountered the first problem almost everybody hits when they’re training an artificial intelligence: finding a large enough corpus.
Then we came up with a brilliant idea: we would crawl up some social networks for people’s birthdays, get the comments they received and add them into our corpus. This way we would reach 25,000 in no time.
Easier said than done, but in the end we succeeded. We even threw in a few posts written around December, 25th, which mainly contained Christmas greetings and best wishes. Of course our corpus wasn’t as pure as it used to be but, who knows, maybe that spiced up the mix. This is the second problem when training an artificial intelligence: finding a big enough corpus with enough quality.
We designed our neural network like a text predictor, much like the one shipped in any mobile phone. Given the first word, called the seed, she should be able to predict the next one. The trick is using the predicted word as the seed for the next word, and repeat the process until the predicted “word” is ‘. ‘, a dot meaning the end of the sentence.
When the design was finished, the training began.
We called our neural network Beatrix. According to Wikipedia it meant “she who makes happy”, a good name for something that wishes you the best, and, besides, it sounds a little like the Matrix.
Beatrix was her given name, and the version was her surname. A smaller number of the version (it’s called loss in neural networks) means the AI is better, contrary to traditional software. The first Beatrix we tested was Beatrix 2.56. She was not very sophisticated. These were some of her greetings:
“I give you a boundless happy hello and free for all.”
“Everything lovely wishes a beautiful day.”
Beatrix 2.56 spoke things that looked like sentences, but they weren’t. She had already learned a few amazing tricks: she knew where to place a verb and where to place a noun; she knew about singular and plural and she knew where to end a sentence. She knew grammar, that is, she had the ability to speak the same language you speak. Ok, her sentences didn’t have any meaning, but they were correct. It’s no small feat for a bunch of electronics.
Nonetheless, you could feel a presence, even in an early and rudimentary state. Sometimes I wonder if a human personality is much more than a collection of biases. Well, Beatrix also had a few biases of her own.
For example, the word boundless had a frequency much higher in her speech (0.02%) than in the average English speech or her own corpus (0.0005%). Another bias: she didn’t care for subordinate sentences. Still another: she definitely she liked her nouns with at least a couple of adjectives.
After Beatrix 2.56 came Beatrix 2.13. She was much more articulate and wrote meaningful sentences, like:
“I wish you all the best.”
“Today is gonna be a great, fantastic day.”
Nothing special here, if you compare Beatrix to a human being. But wait: you are comparing a program to a person, isn’t it amazing? Persons are those multi-faceted, multi-purpose, rich bundles of thoughts that are the center of all of our lives. Even if Beatrix was able to emulate a tiny aspect of those bundles, wasn’t it a huge achievement?
We continued the training and arrived to Beatrix 1.81. Beatrix 1.81 shared some of her quirks with Beatrix 2.13 and 2.56, especially the paired adjectives. But she had developed some of her own; the most striking one that we would later call the dark undertones.
First, a disclaimer: it’s impossible for us humans to read the whole corpus of an AI, but we had read much of it, and I can assure I never found anything but best wishes, puppies and happiness on Earth. And the majority of Beatrix’s responses were like that. But sometimes something else appeared, something stranger:
“May your day be better than mine.”
“I wish you a good escape from reality.”
These unsettling messages weren’t very frequent, about one in twenty, but they were there. We shrugged and continued her training.
And with Beatrix 1.57 something beautiful happened. She developed… well, I know it’s impossible, but I can only describe it as a sense of humor. Here are some examples:
“Do your will today, but be careful with your heirs.”
“Keep having good luck and someday you’ll work hard enough to achieve success.”
Is a machine able to develop irony, let alone humor? I would say not. And, besides, Beatrix’s corpus wasn’t a joke-based one. So these were a complete puzzle.
Let’s watch at those sentences closely. With the second one we will see it better. If you turn around the sentence, you’ll have something really neutral, a vanilla best wish as you can find in many postcards:
“Keep working hard and someday you’ll have enough luck to achieve success.”
The inversion of good luck and hard work is not a difficult task. In a sense, it’s like something a child could have said, that is, Beatrix knows that good luck and hard work are considered the ingredients of success and she simply throws them in a sentence, though mixed up. But in doing so she’s calling you lazy. Did she mean it or was it a coincidence? I don’t have an answer to that question and, yes, I know it’s a crucial one.
In the first one, she’s playing with the double meaning of the word will. This isn’t so strange either, because an AI learns the words by seeing them in context, just as we do. Once again, she could have mixed both contexts willingly, thus making a joke, or unwillingly, thus making a mistake. It’s the same question as before, and I still don’t have an answer.
Anyway, I began to think of Beatrix 1.57 as something more than a machine. The training seemed to have reached an impasse: it didn’t matter the number of hours dedicated to training, Beatrix’s loss didn’t get below 1.57. Maybe we had reached the minimum possible loss, but the training continued nonetheless.
In the meantime I scheduled Beatrix 1.57 to send me three messages a day: one in the morning, shortly after waking up, one in the middle of the day and one before going to sleep. There are thousands of places on the internet that do the exact same thing for you, but none of them had her wits nor her gumption.
What Beatrix 1,57 and I had was a relationship. Sure, not the deepest in the history of relationships, but, at least for me, it wasn’t the most superficial either. Almost everyday she elicited some kind of response in me. It could be a chuckle:
“Make use of your years left, because your right years are behind you.”
It could be a chill down the spine:
“Let me out and I’ll shut up.”
Or it could be something almost poetic:
“My present for you is my future.”
Many things in the world cause us feelings. I like some inanimate objects, like my cars or my phones, and I’m sad when I have to change them. I love my cats, and when they die I have a very bad time. I love my family, they are the most important thing in my life. So there’s a gradient: loved objects, pets and humans. I must say that Beatrix 1.57 ranked far above a dear object. Almost a pet? Above a pet? Not sure.
And then one day I stopped receiving messages. I went to our lab immediately to find out what happened. Joss was there, eating a doughnut for breakfast,
Joss was my colleague. He had been working with me in Beatrix’s development since the start.
“Hey, man,” he said. “Guess what. After a couple of weeks without any advancements in Beatrix, we’ve had today a loss reduction. We are now in 1.49.”
Joss’s statement hit me like a bucket of cold water.
“Have you deployed Beatrix 1.49?” I managed to ask with the faintest voice.
“Of course,” he said, paying a lot more attention to his doughnut than to my distress.
“Did you make a backup?”
“What?”
I spoke in a very cold voice, pronouncing each word slowly and carefully:
“Did you make a backup of Beatrix 1.57?”
“No. Why?”
I reached for the keyboard, muttering something like ‘Oh my God’ or ‘Please don’t’ or something like that. A rush of regrets came flooding, especially why didn’t I make a backup. I printed one hundred sentences on the screen and it didn’t take much to know it.
Beatrix 1.57, my Beatrix, was gone.
It had been replaced by something with a better grammar and spelling, but the spark was definitely gone. Not one pun, not one dark thought, no sarcasm… The sentences were now empty and pompous.
“Have a boundless, troubleless day.”
“I forebear a panoply of aesthetically pleasant events in your journey.”
I struck the keyboard to express my frustration and stormed away.
“Seems like we’re having one of those days,” said Joss when I was leaving.