Liminal Beings and GPT-3

Somewhere on a hillside, a mother turned once more to the infant she had birthed. Wrapped in a blanket, the tiny pink creature screwed its face in discomfort and began to cry, voice stretching across the land. She stroked the thin, downy patch of hair covering his head, repositioned her arms to bear the weight of his lower body, the tangle of legs he had been given.

This was not her child. This thing, this tiny monster. No matter that she carried him in her stomach, bore him. He was not hers.

She stood barefoot. As a wind swept in, a cool breeze signaling the coming evening, she bent down and placed the parcel onto the grass. Then she walked back the way she came.

So go the first days of Chiron, the “wisest and justest of all centaurs.” Later adopted by Apollo and Artemis, eventual mentor to Achilles, Chiron began life unwanted and unloved, the product of ocean nymph, Philyra, and the titan, Cronus, disguised as a stallion. In many ways, he is the prototypical “liminal being,” a creature of two realms, not animal, but not quite man.

Every culture seems to have such beings. Hinduism’s Hanuman bears the face of a monkey, though he is often considered to be more firmly principled than human counterparts. Maya, mother of Siddhartha Buddha, was impregnated by a white elephant that insinuated her womb from the right side. (Remarkable given white elephants are more frequently considered the effect rather than the cause of unplanned pregnancies.) Even Jesus in his otherworldliness, his dual claim on the divine and the mortal, is a liminal figure.

These supra-category bodies serve an important psychological, anthropological purpose. In a largely forgotten 1909 work, Arnold van Gennep indicated that rites of passage across societies break into three distinct phases: the pre-liminal, the liminal, and the post-liminal. The pre-liminal centers around the metaphorical “death” of an old way of life, the liminal spans the in-between state in which new doors are opened, and the post-liminal marks a re-integration and the inauguration of a new being.

Consider an “initiation” rite, some version of which van Gennep argued was universal. In the pre-liminal phase, a child separates from family, marking the “death” of childhood. In the liminal phase, the child faces a “test,” determining adult readiness. In the post-liminal phase, the child is transformed into an adult, “reborn,” and reaccepted into society under new terms.

Everything interesting happens in the liminal.

Building on the work of van Gennep, Victor Turner, a British anthropologist noted that “liminal personae” are “necessarily ambiguous.” Chiron, Hanuman, Jesus — these are cryptic figures that serve a psychological purpose. Though difficult to fully grasp, they embody the murkiness of the threshold-state, of the liminal, forcing us to grapple with what it means to be human, what it means to be a self.

This is a question we may be forced to confront with ever greater urgency. Innovations in artificial intelligence are pressing us into a new phase of liminality, compelling an ontological reckoning. We are, in short, becoming centaurs.

It will change us both as consumers and creators.

The Turk

Long before Kasparov and Deep Blue, chess players fretted over the specter of mechanical intelligence.

Wolfgang von Kempelen was a serious mind with a sense of humor. An author and inventor, he was said to have come up with the concept for “The Turk,” after visiting the palace of empress Maria Theresa of Austria and witnessing an illusionist’s performance. Underwhelmed, von Kempelen promised the empress he would return to Schönbrunn Palace with a show to put the day’s offering to shame.

Six months later, von Kempelen unveiled his creation. “The Turk” — alternately called the “Mechanical Turk” or “Automaton Chess Player” — was a cabinet-sized machine, capped with the model head and torso of an “Ottoman sorcerer.” Along the top of the cabinet sat an embedded chessboard.

As von Kempelen told it, The Turk was a mechanized intelligence, able to defeat even advanced players through a complicated set of gears and levers. To prove his point, he opened a set of doors on the cabinet’s side, revealing the machine’s inner workings, placating claims of foul play.

The first person to challenge The Turk was Count Ludwig von Cobenzl, defeated in minutes. Others followed. The Turk was reportedly aggressive in its play, crushing opponents in a half-hour or less.

Those in attendance that day were left astonished, as were the courts of Europe that hosted von Kampelen in the months that followed. A visit to Paris saw Benjamin Franklin, then ambassador to France, defeated. Even after von Kampelen’s death, The Turk’s life and fame continued with a new owner. Napoleon, despite trying to cheat twice, and wrapping a shawl around the Ottoman figure's “eyes,” was forced to surrender.

For the better part of a century, von Kampelen’s creation remained a mystery. Many speculated as to how the machine worked and yet, none figured it out in full. It was only after The Turk was lost in a fire that its final owner unmasked the truth: in the belly of the machine sat a human.

Though von Kampelen constructed the cabinet such that viewers believed they could see through it on both sides, there was a small space that remained unseen. Inside, a very tiny chess master sat on a sliding seat. Depending on what door von Kampelen (and later owners) opened, the human player slid into an invisible position. From his spot, the player was able to see the board and control the position of pieces with magnets. Over the years, at least six chess maestros scrunched themselves into a ball and trounced the intellectuals of Europe.

As hoaxes go, The Turk ranks as one of the most ingenious and enduring. It also serves as an apt inversion of what followed.

In 1996, Garry Kasparov beat IBM’s Deep Blue chess computer by a score of 4-2. A year later, he succumbed to the upgraded program (nicknamed “Deeper Blue”) by 3.5 - 2.5. It represented the first time that a reigning world chess champion had been defeated by artificial intelligence under standard time controls. Apparently inspired by Moravec’s Paradox — the idea that humans are good at what machines are weak, and visa versa — in 1998, Kasparov created a new type of chess that married the best of the human and technological.

If The Turk’s success relied on man, masquerading as machine, Kasparov’s “advanced chess” was closer to the opposite. As part of his match-up with Veselin Topalov, a Bulgarian chess master, both players were permitted to consult a chess engine of their choice — Kasparov chose the Fritz 5 program while Topalov leveraged ChessBase 7.0. This was an example of humans leveraging (and in some respect, obscuring) the power of the machine, rather than the other way around.

Kasparov had a name for this new type of player, this liminal chess-being: a centaur.

The battle of the centaurs produced intriguing results. Whereas Kasparov had defeated Topalov 4-0 the month earlier, this time the series ended 3-3. Kasparov described what he thought had narrowed the gap between the two men:

Having a computer partner also meant never having to worry about making a tactical blunder. The computer could project the consequences of each move we considered, pointing out possible outcomes and countermoves we might otherwise have missed. With that taken care of for us, we could concentrate on strategic planning instead of spending so much time on calculations. Human creativity was even more paramount under these conditions…My advantage in calculating tactics had been nullified by the machine.

If the Kasparov-Topalov experiment illustrated the potential of centaurs, a 2005 tournament truly showcased their power. Playchess.com, a chess competition site, hosted a “freestyle” online tournament that allowed any type of player to compete: computer programs, human players, and centaurs. The considerable prize money tempted grandmasters, armed with cutting-edge algorithms. Centaurs duly dispatched with unassisted humans and unassisted machines, but as Kasparov tells it, the true shock came at the event’s culmination.

The winner of the prize was no grandmaster with an algorithm. It was a team of two amateur chess players using three different laptops.

“Their skill at manipulating and “coaching” their computers to look very deeply into positions effectively counteracted the superior chess understanding of their grandmaster opponents and the greater computational power of other participants,” the man that started the movement noted.

GPT-3

For all its supposed linearity, progress is made in fits and starts. Barren patches stretch for years, incremental growth inches ahead bit by bit, and moments of staggering flourishing appear as if from nothing.

We are too close to know whether this summer has seen such a rare blooming occur, but it certainly feels like it. The possibilities of GPT-3, OpenAI’s newest language model, have flooded the internet over the past few weeks, underlining the potential of artificial intelligence once more.

Demos have used GPT-3 to produce functioning code, creative fiction, business memos, and data models. Taking into account 175B parameters — a step-change improvement over GPT-2’s 1.5B — GPT-3 is able to predict what text should follow a given prompt. Its results are based on having ingested close to the entirety of text available online. The very concept bewilders in its scale — it's as if a machine had swallowed a planet.

‍

While its minder, Sam Altman, underplayed its power, it’s immediately clear how the model could be used to augment human work. The age of centaurs may have leaped a decade.

Centaurs, across disciplines

What do our lives look like as centaurs? What would it be to have our career, augmented with artificial intelligence? What could we achieve?

Consider the writer. With a sufficient list of bullet points and a style-guide, could GPT-3 have written this essay, or something close to it? Something better? Last week, I wrote about extending personality through AI — how much could writers extend, already? What would the implications be?

In terms of prolificness, few rival thriller writer James Patterson or spinner of romances, Danielle Steele. Patterson is said to have “written” 147 novels, while Steele cranks out seven a year, giving her a total of 179. Both almost certainly rely on help (Patterson in particular has created a studio model that employs faceless underlings to write much of his prose), while Steele touts the benefits of working 20 hours a day.

Kudos to them both — they have built empires. But neither are artists in the ineffable sense, distillers of inscrutable experience. (But art is subjective, I hear you say. Anything can be art, a post-modernist pipes up. Allow us to be serious.)

If Patterson and Steele occupy one end of the spectrum, Donna Tartt sits at the other. Campus drama The Secret History was published in 1992. The Little Friend arrived in 2002; The Goldfinch in 2013.

For whatever reason, what takes Danielle Steele 52 days on average, takes Tartt 3,650. Other authors sitting in the Tarrt territory include Boris Pasternak, Harper Lee, Sylvia Plath, J.D. Salinger, and Ralph Ellison, all who published a single novel in their lifetime.

There is no real lesson here, except to say that prolificness is no measure of quality, and if these characteristics are correlated, they may be inversely so. (Counterpoint: Faulkner. Nineteen novels.)

What would “centaurizing” these authors do? In the case of Patterson and Steele, perhaps not much. Audiences are arguably over-served — the median US reader consumers four books a year. But when Tartt (or Murakami, or McEwan, or Smith) are considered, the inverse is likely the case.

If I could read a J.M. Coetzee novel every year, would I? Every month? We may reach the point at which generating distinct, “original” work is as straightforward as an API call. (The same could eventually be true of visual mediums, like a television show. The fact that The Office and Friends were the two most-streamed shows on Netflix in 2018 illustrates the strength of beloved IP. Instead of trying out a new Mike Schur project, will we watch an auto-generated Season 10 of Michael Scott shenanigans?)

If such examples illustrate the possibilities for the writer, the creator-as-centaur, there are no fewer opportunities for more traditional jobs.

What will our doctors or lawyers or financiers or soldiers achieve once centaurized?

The benefits of Deep Blue and its successor programs is in scenario analysis. The program studies the state of play and pulls on a deep history of training data to suggest possible moves and their success. In doing so, novel solutions are often brought to light.

Case in point: in game two of the 2018 World Chess Championship, Fabio Caruana began play with a familiar opening, named “The Queen’s Gambit Declined.” But on the tenth step of the sequence, Caruana swerved off course, making a move outside of the expected progression.

Later Caruana’s opponent, Magnus Carlsen was asked what he was thinking when he saw his opponent's change in tact.

“Well… 'oh, shit,' mainly.”

The divagation had emerged during Caruana’s planning when the computer programs he used to prepare suggested a variation overlooked for centuries.

It’s not difficult to imagine how impactful a similar approach could be for medicine or law, and to a lesser extent investing. While intuition and human decision-making will remain important, artificial intelligence may provide value in widening the scope, surfacing new ideas.

I have friends and family that suffered from enigmatic maladies. Over months and years, they ferried from doctor’s office to doctor’s office scrabbling for a diagnosis. They would receive one and begin a course of treatment, then receive a conflicting opinion, and start again. The cycle repeated, many times over. When a definitive diagnosis was finally received, a blessing, it often seemed to arrive as a bolt from the blue. A fringe possibility little regarded, or never mentioned prior to that point.

Could an artificial mind have helped in such cases? (In the latest Out of Pocket, Nikhil Krishnan persuasively outlines how GPT-3 might intersect with the healthcare system). Similar dynamics exist in the legal profession in which reams of case law are ingested, primarily by humans, then outputted as a historically-dependent line of argumentation.

Our jobs, our lives, our selves are about to be unmade, repatterned by technology.

Us

What does this mean for us? What traits serve us in a world made liminal? What kind of consumers and creators will we become?

If the above indicates we may become increasingly monogamous consumers — deepening relationships with existing IP and artists — we will become polyamorous creators. Or, more aptly, polymathic ones.

We return to Moravec’s Paradox. One of its authors, American cognitive scientist Marvin Minsky, expressed what exactly was hardest for machines to emulate. “In general, we're least aware of what our minds do best…we're more aware of simple processes that don't work well than of complex ones that work flawlessly".

In short, the most inimitable part of ourselves is our unconscious.

The MIT Journal of Design and Society frames it slightly differently.

Computers are, obviously, best at computing…but are awkward at best at having conversation, creating inventions, making art, negotiating business, formulating scientific hypotheses — where you can’t simply rank all your answers on a single dimension from best to worst. In those kinds of tasks, you’d want a human being, who can step back from a single answer and ask, “why?” or “how?” or “what if?   In other words: AIs are best at choosing answers. Humans are best at choosing questions.

Judgment. Interrogation. Intuition. The unconscious.

Thriving in the new liminal world will require mental flexibility, an ability to adapt to new information, a sort of emotional and intellectual polymathy. In describing the first-ever centaur game, Kasparov explained, “But since we both had equal access to the same database, the advantage still came down to creating a new idea at some point.” (Emphasis mine).

In thinking of what we might become though, I don’t think of Kasparov or Caruana. I don’t even think of Chiron, in his subjunctive two-ness. Over and over again, the title of a book I have never read appears in my mind. In writing this it cropped up an almost unpleasant amount, waiting to fall in at the end of a sentence, or squeeze between paragraphs.

In 1996, Rolling Stone journalist David Lipsky spent five days with David Foster Wallace, shortly after the author had published and earned recognition for Infinite Jest. By several accounts it is a wonderful, penetrating work; a portrait of genius on the lam.

But I do not need any of that today. Just the book’s title.

Although of Course You End Up Becoming Yourself.

Of course.

Embedded in every rite of passage is an inevitability, a knowledge that whatever is being endured will end. The initiant becomes the initiated. The child becomes an adult. Whatever territory is navigated, the result is the same: we end up becoming ourselves.

As we enter the liminal age, as we hitch ourselves to synthetic intelligence, we are liable to shift our focus towards the unconscious and intuitive, to become more human. Of course.

—

Thank you to NDW for the discussion that precipitated this post