Human, All Too Human 1
Really?
I meant to avoid writing about popular media topics when I started this newsletter. I thought I would do the opposite and write mainly about truly important things like my mother’s antique furniture. But I couldn’t resist fiddling with ChatGPT, which, after all, is supposed to be able to emulate human thought and is an existential threat to us all, especially anyone who makes a living as a writer, erm, content producer.
I can read the headlines as well as the next person, and I am aware that the end of the world as we know it is at hand. (Not a bad thing actually, in my view.)
Beware, I say, the Chattering Machines, for they shall inherit the earth.
Okay, but really? The Luddite position here is always retrograde (and boring). There is absolutely no point in crying doom about a new gadget until you have messed with it a while, invited it into your home, taken it out for a beer, and seen what it looks like first thing in the morning before coffee. I set up an account at ChatGPT and started conversing with it, you know, out of curiosity. Well, curiosity tempered with a certain amount of mischievous disbelief. (Whenever I read one of those headlines, I say, Really?)
How it works
Before that, a little summary of my reading about ChatGPT. First, it’s an acronym for Chat Generative Pre-trained Transformer. Right away, not so interesting. It was originated by a man named Sam Altman from St. Louis, Missouri. He is now the CEO of the parent company. The model has been through, I think, three iterations. We’re now on GPT 3. (GPT 4 has just come out, but I am not sure exactly when or if I have met it yet.) My son the programmer downloaded GPT 1 when he was an undergrad at the University of Waterloo, trained it with a database of downloaded criminal investigation interviews, and had very creepy late night conversations with it, which he would then relay to me via email, so I could put them in the novel I was working. (So there you go. Already, ten years ago I was incorporating GPT into my fiction. They don’t call me Cutting-Edge Glover for nothing.)
The idea or pre-history of Artificial Intelligence is not old. Basically, it dates back to a 1950 essay by Alan Turing (a brave Englishman who contributed greatly to the effort to defeat Germany by helping decode the Enigma machine and then was hounded for his homosexuality until he committed suicide). The essay was called “Computing Machinery and Intelligence" and was published in a British philosophical journal called Mind. Turing said that if you keep asking can machines think, you’re asking the wrong question. He thought it better to ask whether a machine can be made to imitate what humans do when they think they are thinking. That is, can you invent a machine that outputs the same sort of thing that humans produce and indicates that they are thinking? In other words, you can bypass “thinking” altogether if you can just produce outputs that imitate human outputs.
It turns out that actually it’s not too difficult to imitate human outputs. Machine learning programs are essentially predictive devices set up to “learn” from a large number of examples. In the case of something like ChatGPT, the learning is accomplished by feeding the program large amounts of published text (the exact texts and amounts are not public knowledge). The program doesn’t know what it’s been reading in the human sense. It assigns a number to each particular word (or word fragments or sometimes whole phrases) and then assembles these into something like clusters of numbers (words) that often appear together. This allows the program to give a rough prediction of what word (number) should appear next in any sequence. To make this process slightly less mechanical and more “creative”, the program writers have ways of weighting certain words or groupings and also a way of randomizing predictions (say, if the database offers ChatGPT the ten most often used words in a certain situation, the randomizing factor might instruct the program to pick the 18th most often used alternative). This inbuilt unpredictability makes the program look even more human.
Computer programs can now defeat Grand Masters at chess and write undergraduate English essays, not to mention do surgery and solve very complex mathematical problems. It took just over 50 years from Turing’s paper for this to happen. And if you still want to ask, well, is the machine thinking, you might want to turn the question around and ask yourselves what is it you are doing when you are thinking? It’s a very difficult question to answer.
One thought: Maybe we have been kidding ourselves all along. Maybe humans aren’t that bright.
My first go-round with ChatGPT
But enough introductory (I had much more to say about machine learning and large language models, but, really, you can just go ask ChatGPT to explain it).
My first go-round with ChatGPT went like this. I quote without editing.
DG: Write me a short story about a girl named Dawn and her dog Rusty and her boyfriend Loy Garage in the style of James Joyce.
Bot: Dawn and her dog Rusty were strolling along the seashore, the salty breeze tousling her golden hair and Rusty barking joyfully at the waves. She thought of her boyfriend, Loy Garage, and how he would love to join them on these walks. Loy was always working at the garage, fixing cars and dreaming of opening his own repair shop one day.
As they walked, Rusty sniffed at something in the sand and started digging frantically. Dawn peered over to see what had caught his attention, and her heart leapt as she saw a small box buried there. She dug it up and opened it, revealing a beautiful gold ring and a note from Loy. The note read:
"My dearest Dawn, I may not have much to offer now, but I promise to love and cherish you always. Will you marry me?"
Tears streamed down her face as she slipped the ring on her finger and hugged Rusty tightly. She ran back to the garage, eager to say yes to the man she loved. Loy was waiting for her, a bouquet of wildflowers in hand, and they embraced as Rusty barked and jumped with delight.
From that day on, Dawn and Loy were inseparable, spending every moment they could together, with Rusty by their side. As Loy's garage became more successful, they travelled the world, always accompanied by their faithful companion. And though life was never easy, they faced every challenge hand in hand, knowing that their love was all they needed to see them through.
This was surprisingly good, I thought, but ultimately disappointing, given the initial assignment and my expectation that the bot had read Joyce and should have produced a better class of simulation. That is, it might be emulating human thought, but for the most part not the kind of human I’d want to sit over a beer with.
I was brutally frank in my criticism. (My former students will remember getting responses like this.)
DG: That doesn't sound like James Joyce.
Bot: I apologize if the short story did not meet your expectations in terms of style. James Joyce was known for his modernist writing style, characterized by fragmented narratives, stream-of-consciousness, and experimentation with language. Here is a revised version that incorporates some of these elements:
Dawn and Rusty, her dog, walked the seashore. The wind tore at her hair, while Rusty barked at the waves. She thought of Loy, at the garage as always, fixing cars, dreaming of his own shop. She remembered his words: "I may not have much, but I love you."
Rusty dug in the sand, revealing a box. In it, a ring and a note from Loy: "Marry me."
Tears, laughter, a rush to the garage. Loy, a bouquet, waiting. Embraces, barks, a life together. Cars, travel, the world. Challenges, but love enough. Always Rusty, always Loy, always Dawn.
Together, hand in hand, heart in heart, they wandered the shore, the future unknown, the past forgotten.
At this stage, I was deeply impressed but uneasy. I was impressed because I had not expected anything remotely this competent. The bot wrote these texts in about 30 seconds. They are grammatical and actually narrate little stories, present scenes, and give thoughts and dialogue to the characters. I loved the dog digging in the sand and finding the ring. Hell I loved this:
Tears, laughter, a rush to the garage. Loy, a bouquet, waiting. Embraces, barks, a life together. Cars, travel, the world. Challenges, but love enough. Always Rusty, always Loy, always Dawn.
Best of all the machine apologized — so gracefully, so immediately. Seeing this, I felt all sorts of resistances crumbling inside me. I was almost won over. No student in my years of teaching had ever been less emotionally needy and so easily and productively influenced by my suggestions.
On the other hand, it was not Joyce. Not even close. And the mawkish sentimentality was cloying — to me; I realize a lot of human readers would eat this up. In a sense, my first discovery was that the machine might actually be too human to be interesting.
Anyway, I pushed on. I wanted to challenge the program. Note again, how bloody affable ChatGPT is.
DG: Can you write the same story in the style of Herman Melville?
Bot: Certainly, here is the revised story in the style of Herman Melville:
Amidst the rolling breakers of the ocean, Dawn and her canine companion Rusty roamed the sands. The tempest of the wind howled through Dawn's tresses, as Rusty barked his joy at the tumultuous sea. In her thoughts, she pondered upon her paramour, Loy, who worked ceaselessly at his garage, ever toiling with engines, ever envisioning his own repair shop.
Rusty sniffed at a buried object in the sand, and began to dig. Dawn peered over, and discovered a small casket, within which lay a golden ring and a missive from Loy. The inscription read:
"My beloved Dawn, I am but a humble mechanic, yet I vow to cherish you forevermore. Will you be my bride?"
Tears coursed down her cheeks, as she placed the ring upon her finger and clasped Rusty close. She hastened back to the garage, eager to plight her troth to the man she adored. Loy awaited her, bearing a bouquet of salt-tossed flowers, and they enfolded each other, Rusty barking in gleeful unison.
From that day forth, Dawn and Loy were bound together, ever seeking new adventures, with Rusty their steadfast ally. Though life's trials were many, their love was a beacon, guiding them through storm and calm alike. And so, they journeyed on, a trinity of love and devotion, always Rusty, always Loy, always Dawn.
This one is even better. Third try is a charm. The language is, yes, ever so slightly more Melvillean than Joycean (those rolling breakers, so characteristic of the sailor-turned-author). And there is that lovely phrase “salt-tossed flowers.” Good Lord. The sentences are more complex. But underneath, the same mawkish sentimentality.
My thought at this stage is that much of what goes by the name of culture and literary art is more cliché than we normally think. As I say, engaging with chat bots should make you look at yourself and less at the bot, which, after all is only a predictive machine crunching numbers (it predicts sequences of words not meaning). If these little texts look like short stories (at all) it is a judgement on what we normally think of as short stories.
Most of what goes by the name of literature is human-generated machine prose. I call it novelese, by which I mean a certain very common kind of writing that merely strings together sentences, characters, scenes, and dialogue that look creative, look like novels (hell, it’s 300 pages long, must be a novel) but really aren’t worth reading. They are very sophisticated clichés. Publishers like these kinds of works because they are easy to commodify and lots of readers buy them because they have learned (been programmed) not to expect any more.
In other words, what I learned today was not that chat bots produce clichés, but that humans produce a lot of writing that looks like it could have been written by a bot.
In other words, Artificial Intelligence is US.
I shall return in a week with the next chapter of The Battle of the Bots — DG Bot Vs. ChatGPT.