
With Centaur Books, I am not referring to the legendary New York fantasy book publisher extinct in 1981. No, I am talking of an entirely new literary genre, the creative interplay of humans and machines, illustrated by the mythical creature combining man and animal.
Books by AI have been around for several years but very few people care to read them. The best of the category, such as 1 The Road, a project curated by Ross Goodwin and Kenric McDowell, can be weirdly beautiful but too experimental to appeal to the larger book public. In contrast, Fear is the Key, under my pen name Lars Amber, **** is the first full-length and "understandable" novel written with the help of an advanced text-generating model. Its narrative is coherent from the first paragraph to the last. The thriller author as a curator! This story is about the making of it.
To start, we have to travel a long way back, to the last century. Gary Kasparov, the world chess champion, was beaten by IBM´s Deep Blue. The year was 1997. Gary is still active as an elite player though, informally ranked as number five. AI is not his sworn enemy, on the contrary. The year after his historic defeat, the Russian grandmaster introduced a new way of playing which he named Centaur Chess. In 2005, then, a Centaur chess tournament demonstrated that a human player assisted by AI may win over the best AI chess machine solo. A creative concept made thus its way to other AI areas. Humans are better at imagining new paths, whereas machines are unbeatable in solving problems with known structures, relying on stored memory and massive data.
Meanwhile, the issue may be more complex. The Chinese board game of GO is seen as the most strategic, with a much higher number of possible positions than in chess. In 2016, Google´s AI machine AlphaGo did beat the best human GO player with a bizarre move that no human would have thought of. In the absence of a better definition to describe the dynamic of layers upon layers of self-improving neural networks, one might talk of an unexplainable intuition hidden among some of the best Algorithms. Today´s artificial neural networks are indeed elusive in their ways of assembling words and even more so contexts. Cognitive scientists fail to explain how they really function, step by step. ANNs sometimes produce poetry without being poets themselves, small pearls of beautifully crafted sentences in the middle of pure gibberish.

"I have applied Kasparov´s centaur concept to literature."
An AI machine will never write like Marcel Proust, the acclaimed author of In Search of Lost Time. That is, not before we have invented artificial general intelligence, the much-coveted AGI! It is through language that the human brain can reach some of its most sophisticated performances, especially due to its capacity of using symbols and expressing emotions. The day a computer writes a good novel from A to Z by itself, it has already acquired a superior human intelligence and will likely be on its way to achieve supra-human intellectual tasks in all fields.
In the meantime, AI programs may become powerful creative teammates. I have applied Kasparov´s centaur concept to literature. Even if Fear is the Key is a pilot novel, the fruit of the Covid-19 pandemic, I bet that the genre will become mainstream soon.
In 2016, Japanese researchers at the Kimagure Artificial Intelligence Writer Project published The Day a Computer Wrote a Novel. Their purpose was to uncover the very nature of creativity. It was a success that was hailed by a unanimous international trade press for its stylistic qualities (it was almost given a literary award!). The Japanese "novel" has a length of 10,000 characters though, which per definition places it among short stories (a very short one). Not more than twenty percent of its content is AI-generated, according to a participant in the project group. Fear is the Key, on the other hand, has a length of 78,000 words. For a techno-literary comparison, Casino Royale, Ian Fleming´s first novel introducing the character James Bond, counts 65,000 words. My objective was to reach fifty percent of AI-content. With that, I was not successful, but I think that my score of 37 percent (circa) is pretty decent. If I include inspirations given by the Machine, 50 percent of this book´s content is not due to my creativity alone. I, therefore, believe that Fear is the Key is unique.
"Text robots are now creative"
This is a fact: we should not any longer consider that creativity in writing is solely the prerogative of humans. Maybe it is by accident, but text robots are now creative. The Day a Computer Wrote a Novel hinted that position five years ago. Fear is the Key may be seen as statistical proof of concept.
The field of Natural Language Generation (NLG, i.e. computer-generated text) is speeding up nearly as fast as light. I have been using the GPT-2 language program, already obsolete. It was developed by OpenAI in 2019 (co-founded by Elon Musk in 2016 who exited the company´s board three years later though), today with a hybrid status of a for-profit and a non-profit.
More specifically I made use of the 1.5B GPT-2 model, the most complete in this generation. As I am not a techie myself, I got precious assistance from several people and organizations in the operational phases. I first discussed the project with Erick Fonseca, a Natural Language Processing academic (NLP, i.e. what occurs when machines read language), and a blogger from Brazil who is presently doing a Postdoc at the Instituto de Telecommunicações in Lisbon. Erick volunteered to partner. Because of a tight planning schedule for him in Portugal (and lack of some adequate hardware), I had to find another technical support, which led me to the ancient Indian city of Ahmedabad, a fast-growing services hub. There, the 1.5B model was fine-tuned by Pragnakalp Techlabs and its founder Mittal Patel. It turned out to be not so trivial. It only worked after backing from a monster Google Cloud TPU (Tensor Processing Unit). Thus, some 600,000 words were eventually generated. NLP researchers call those "tokens". No matter the denomination, the output corresponds to six large novels.
That output had been activated by hundreds of Writing Prompts formulated by me. Your humble servant also selected the content of the 16 MB Corpus of Literature dataset used for the fine-tuning. It consists of some 35 novels and many minor documents. Each chapter (forty-four in all, plus a prologue) starts with an epigraph, a literary quote, most of them taken from the Corpus.
"At times I felt like a cowboy trying not to fall from the saddle in a wild horse rodeo!"
Now we live in a world of insanity Fear is the key to what you want to be You don’t get a say the majority gets its way You’re outnumbered by the bastards till the day you die
The main epigraph at the beginning of the novel is taken from an Iron Maiden song. It contains my very book title which was internet-researched to reflect the permanent state of anxiety which is dominating our present living days, more than ever before. In the same spirit, the visual elements composing the book cover´s collage are internet-search-generated. Conspiracy = Templar ; Action = Agent ; Future = Quantum computing; Manuscript = Handwritten; Time = principal variable of a never-ending Algorithm… But the cover artwork, thought of like a Dan Brown pastiche, is entirely human, created by the graphic designer Jan Felipe Beer (though obviously also on a computer!).
We should mention the project´s drawbacks. Text-generating models do not – yet – constitute a working tool for delivering long texts with a thorough sense. It took me more time and rendered more stress to produce Fear is the Key in this machine-way than previous literary works in a human context only. At times I felt like a cowboy trying not to fall from the saddle in a wild horse rodeo! It is still a painful struggle. I have the intuition that one of the problems of the GPT-2 is the poor quality of many of the eight million web pages on which it was originally pre-trained. To a large extent, these dataset-collected pages were taken from Reddit posts in discussion groups. They often reflect biases and prejudices of a, for me, shocking dimension. Add to that the natural disposition of GPT-2 in rapidly switching into mumbo-jumbo mode, due to its incapacity of appreciating the real world. GPT-2 adores a conspiracy; it has a liking for porn; it seems to detest the European Union without any sound argument to back it up…
But in its rare moments of brightness, it produces sentences or even paragraphs worthy of Jorge Luis Borges. Or Donald E. Westlake…
Here follows an excerpt from the first pages of Fear is the Key:
Fragments of a postscript to the prologue – While a sober analysis might tend to emphasize that, for the first time in the history of humanity, we will be allowed to control the weather and possibly the biosphere through an entirely new category of an algorithm, we cannot ignore that the years and even decades to come will be dramatic. The first assessment obscures the context in which we have chosen to live. In such a context, indeed, it is a major error to be silent. No doubt the French green activist Laure Duhamel and her billionaire uncle believed they could ensure an appropriate answer to the hopes of humanity. From time to time, we come across the usual global megalomaniac financiers, the ones who are to be found in every book dealing with the links between politics, business, and international welfare. This story is no exception.
Except for the named characters, which I inserted, and two or three very minor linguistic adjustments, the whole paragraph above is AI-produced. Here is another:
The place, it seemed to him, possessed a strange consciousness of its own. It conveyed a certain kind of shamanistic energy. The atmosphere and absence of noise made him feel like he was at someplace where artifacts, rather than people, had been essential components of life. The quieter the room, the more homogenous the atmosphere. He recalled the feeling of admiration when he had debarked at La Citadelle´s pier: that of being in the presence of an ongoing masterpiece, an unreal creation, whose qualities were in the grip of unknown forces. The bow windows had museum-class, anti-reflective glass providing clear, unobstructed views while reducing unwanted glare and reflection. Part of them looked directly out on the Gulf but most were directed so the whole landscape of the island looked inviting. The mansion counted other tastefully elegant stained glass windows but vertically designed.
And here comes some cool action at the end of the novel:
Immediately before being hit by the avalanche, he had been enveloped by a sky-high column of white. The snow all around him was blowing like a whirling wind. In seconds he was blinded and tumbled downhill under the chute. He felt the crash of something heavy in his ribs. Well, if this was a decent-sized avalanche, he was bound to be hurt, maybe by his own skis. It felt as if an ice crevice was opening under his right foot and his leg swelled painfully up and down, under the weight of white, toxic slabs. Something wet trickled from the upper section of his right knee. The rest of the leg felt battered. The only significant sound now was the crunch of his heartbeat, in the shallow, looping pattern, and the crack of his ice-encrusted teeth in protest against the weight of the arm. Time and time again he heard the familiar click of snow cracks under his gloved right hand. There was a dark curtain, solid and motionless, surrounding him. He lost all control of his pain in the instant. He had never experienced such fear.
In small chunks, we are right to find this awesome. Even so, I think OpenAI should follow the advice given by my character Dr. Anita Alm, the Swedish Wunderkind of AI research in Fear is the Key, to achieve a better text generation machine through additional open-ended algorithms (to which you will be introduced by reading the novel). But to be honest, I have not yet tried the GPT-3, released in June 2020, which will perform much better than the GPT-2, as it is trained on 175 billion parameters (a hundred times more than its predecessor). Rather strangely, as the entrepreneur and likewise data scientist Julien Lauret points out, "GPT-3 doesn’t learn from experience; it was trained only once and the model weights are not updated when we query it." More powerful but less human, then! But the GPT-3 is not only text-smart, it can also write basic coding and translate to and from English, even if these functions were not intended by its inventors.
For generating a substantial part of my next novel by AI, I need to get invited by OpenAi to launch a fine-tuning project of the GPT-3 (maybe there will also be a GPT-4 ready by then). In my view, this should serve the company´s interests, displaying the text-generation potential to the broader public.
Equally to the American fiction author Robin Sloan, I am convinced that the principal benefit of the most advanced text-generation machines is not their actual verbatim output, but rather the new angles of inspiration they may contribute with. Brave writing colleagues, you will not be out of work in the future! But you may have found a sparring partner of the mind that never sleeps.
To conclude, I must warn the fine connoisseurs of Shakespeare´s idiom. GPT-2 never disposed of a good teacher in English. Most AI fiction texts need to be linguistically revised and edited before publishing. But that also goes for human writers, okay?