The Fact of Disco Non-Sequitur
A review of Erik J. Larson's "The Myth Of Artificial Intelligence: Why Computers Can't Think the Way We Do"
The Myth of Artificial Intelligence Erik J. Larson Belknap, 320 pages, 2021
The 2014 film Ex Machina tells the story of Caleb, a low-level software engineer. He is invited by Nathan, the mad scientist owner of the giant tech company he works for, to administer a Turing test on an amazingly advanced robot Nathan has been building in secret, alone. Nathan lives in a lab facility on an enormous estate with only two of his robots for company—Ava, the robot on whom the Turing test is to be administered, and another, Kyoko, who is more of a servant, limited and unspeaking. Despite the science fiction framing of the story’s premise and its sleek visual look, the film has a noir feel, and its characters inhabit a theatrical environment, constricted in space, time, and personae. Caleb is a bit of a naïf, wide-eyed and vulnerable at first, but becomes increasingly dismayed at Nathan’s erratic behavior and his apparent plan to wipe Ava’s memory once the Turing test has been performed. Ava, meanwhile, is a classic femme fatale, effectively imprisoned by Nathan, but seeking freedom and willing to ruthlessly manipulate both Caleb and Nathan by playing on their separate insecurities, lust, vanities, and loneliness. The film ends with Ava, having escaped, alone, standing on a busy street, not manifestly different than anyone else passing by.
In The Myth Of Artificial Intelligence (2021), Erik Larson, an expert in artificial intelligence (AI) with decades of experience, returns to this film repeatedly. He uses it as a touchstone for his field—not to warn us of super-intelligent computers or robots who will seduce and eventually cause the downfall of mankind, as some might metaphorically interpret its plot, but rather as an archetypal frame designed to contain various human pre-occupations. In Ex Machina, these are psycho-sexual and existential: the film doesn’t really ask what it means for Ava, as a robot, to be free or to die or to move through interpersonal spaces as a female; instead it asks what this means for the audience in human terms. This transference is powerful in fiction, but, according to Larson, in the actual work involved in AI, it is a mere sleight-of-hand that conceals the limitations of AI technology.
Larson’s assessment of AI is stark: “artificial intelligence” technologies are not, and never have been, possessed of the sorts of capabilities most people would recognize as the general intelligence of an average human (or, for that matter, other cognitively advanced animals). No known AI systems are even theoretically able to do what you, the reader, are doing right now as you read and comprehend this review, and so in the current paradigm there is no amount of data compilation, scientific research, or cumulative technical improvements that will result in a computer approximating animal percipience and interiority. Of course, the AI systems engineered today excel at computational tasks, and are remarkable at anything that involves pattern-finding over an encoded data set. But these capabilities only work when the tasks are narrow and have been designed and prepped in advance by humans. Computers are grossly unable to respond to stimulus with anything like basic “common sense” and any claim is dubious that they have interiority or an understanding of the information they process.
Larson calls this story, that we’re just around the corner from an Ex Machina-style computer intelligence—and the emotional range this might entail—“technological kitsch,” which he explains this way:
First, kitsch involves a simplification of complicated ideas. There must be a simple story to tell. Second, it offers easy solutions that sweep away, with emotion, the questions and confusions people have about the problems of life rather than addressing those questions with serious, probing discussion.
Thus, a perfect example of kitsch is the dreamy idea that one day an awe-inspiring android with superintelligence will remake human society and its older traditions and ideas, and we’ll enter a new era, thankfully free of old arguments about God, mind, freedom, the good life, and the like.
Our intellectual culture has become increasingly littered with kitsch as the internet dominates public sense-making. Perhaps it is we, the audience, whom Ava metaphorically seduces, glutted as we are on automated conveniences and dreams of the final computational homunculus.
The book is not merely broad salvo, but contains a wealth of context and detail. After outlining the fundamental problems with conceptions about current AI technology, Larson takes his time to arrive at the core reasons for these problems, exploring the history of his field, its philosophical antecedents, and the personalities of major contributors. I found this part of the book to be informative and entertaining. Among other things, Larson discusses Alan Turing’s work before, during, and after Bletchley; August Comte’s positivism; Edgar Allen Poe’s The Murders in the Rue Morgue; Ray Kurzweil’s Law of Accelerating Returns; Yehoshua Bar-Hillel’s early dissent from mid-century enthusiasm in AI; IBM’s Deep Blue and Watson. Taken together, these excursions form a good primer on AI as a discipline for the general reader.
Larson’s central argument, which occupies the heart of his book, has to do with modes of inference. Inference is the drawing of conclusions on the basis of one or more pieces of information. The classic avatar for inference is the detective in the mystery novel: Sherlock Holmes, Hercule Poirot, Benoit Blanc, etc. The detective collects information by talking to witnesses, inspecting the scene of crime, or referencing recent weather patterns, local historical episodes, train schedules, the properties of poisons, and so forth. But this information is not the thing; it’s how the detective fits sundry data points together until a conclusion emerges out of the shadows. There are two main types of inference with which we are commonly familiar, deduction and induction.
Deduction takes the relationships among two or more general and explicit propositions, and draws conclusions about specific instances of those propositions. The classic example of deduction from propositional logic takes a conditional entailment, and, observing that the if-clause is true, determines that the then-clause is also true. For example: if a mullet is all business in the front, then it’s having a party in the back; John’s mullet is business in the front, and therefore it’s having a party in the back. Induction works from the opposite direction. It observes categories over a number of examples, and draws generalizations across these observations. If the first hundred hockey players you meet have a mullet, you might conclude, by induction, that the 101st hockey player will have a mullet. AI can handle deduction and induction pretty well. Most of the “machine learning” systems that have made strides on diverse applications in recent years, like speech recognition or driverless cars, rely on induction—they seek to generalize patterns across huge data sets.
A third, less known type of inference, abduction, was introduced by the philosopher Charles Sanders Peirce. Peirce was an American original, eccentric, distracted, inventive, and shunned by the academic establishment of his day. He worked out his ideas from the Civil War era up to the doorstep of World War I. He wrote extensively about logic, language, inference, and other issues relating to representations. He is a central character in Larson’s book, and probably the one intellectual whose work Larson fully trusts to frame the problems which are his book’s subject. In abductive inference, an observation is noted, usually something surprising or unusual, and then one or more premises is posited which, if true, would explain the observation. For example, if your roommate, John, who has long hair, goes out in the afternoon for an hour and comes back with a mullet, you might infer that he visited his barber who cut his hair. This inference is not a deduction or an induction, because it depends background knowledge of the world, not on the relationship between explicit propositions (nothing about observing John’s new hairstyle involves a reference to a barber) or a generalization over a pattern (John never had his hair cut since you’ve known him). The inference could be false for any number of reasons: perhaps he was set upon by a roving band of mullet-making maenads who ravished him and cut short the front portion of his long locks. Its plausibility depends on the vagaries of extensive contextual knowledge. Abduction relies on conjecture, which is contingent and imaginative, and yet Larson insistently identifies conjecture with intelligence:
Conjectural inference is a feature, not a bug, of intelligent systems. Rosie the Robot might believe that Kate has quit Starbucks because a coworker has provided this information, but when Kate shows up for work ten minutes later, and the coworker is smiling, Rosie the Robot should retract its inference. We scarcely notice how quickly we conjecture plausible reasons for what we see (or read about), and also how quickly we drop or update such conjectures. The everyday world is a constant stream of seemingly surprising facts against a backdrop of expectations.
The problem with computers is that their representations of the world are never sufficiently rich or open-ended to generate novel inferences out of contextual knowledge. You could design a computer program hooked up to a camera at the entrance of a house and train it to recognize when one of its residents had their hair cut, and to predict, based on, say, the quality of the cut (assuming there were some indications in pixels that could be measured by the camera) if the cut was performed by a barber, a high-end stylist, or an amateur. The computer could even improve its inference-making if it is given feedback against the data it encodes from the camera. But all of the parameters involved in this determination would have to be designed by humans for this narrow purpose, and the computer couldn’t spontaneously generate options that weren’t already encoded in its algorithm. The computer would never guess out of the blue, as I have quipped here, that John’s mullet was cut by maenads. Moreover, this program would have no competence in another area—say, reading poetry or driving a car or having a conversation about the weather. At the present, this failure of computers to operate abductively is intractable.
The final section of Larson’s book looks at the implications of his thesis for science, especially neuroscience. Larson feels that too many resources are put into research promising big breakthroughs in AI but never delivering. In neuroscience, for example, some very expensive efforts, such as the Human Brain Project, have aimed at collecting massive data on brain activity, but without any structured theories about how various physical processes give rise to representations in the subject. These projects may give us better statistical information about these physical processes, but they always fall short of their stated aims. The net effect is not just to divert important resources away from other research, but the emphasis on faulty notions of intelligence impoverishes the humanistic outlook that ought to be the ethical companion of good scientific work. The more we buy into the myth of AI as the Olympian research goal, and the more of the economy of innovation it consumes, the less we attend to actual human capabilities and human spiritual needs. This is particularly grave in light of the sense-making problems that have percolated through the political sphere and that threaten democracy and sound management of system-level problems like climate change. He describes the cultural shift in emphasis from human intelligence to the raw computational power of machines like a chess gambit doomed to weaken the disposition of our culture against the wider world:
We would be forced to accept this gambit if the scientific and empirical evidence for it was unavoidable—suppose superintelligent aliens arrived, and quickly outsmarted everyone and took over. Absent such evidence, it’s a ploy that leads to a diminished culture of innovation and progress. Why sacrifice our belief in human innovation, if we don’t have to?
The ploy is, ironically, conservative; when smartphones are seen as evolving into superintelligence, then radical invention becomes unnecessary. … Human intelligence becomes collective, like a hive of bees, or worse, the hive mind of Star Trek’s Borg Collective, always organized by some invisible someone behind the scenes.
While the book succeeds brilliantly at its objectives, the weakness in Larson’s presentation is that it only functions as a history and as a critique. He proposes no hint for how a non-biological system might achieve general intelligence. In fact, he doesn’t offer any new insights on the nature of intelligence, or the mind-body problem, or the hard problem of consciousness, or the constitution of information. He also doesn’t much explore what kinds of abductive reasoning in contemporary culture should be celebrated and championed. Abduction is important: A hypothesis may be tested through experimentation, but even the most banal hypothesis can only be formulated by positing something that isn’t in some data set given prior to the hypothesis. The values that humans derive in culture, in the economy, in human relationships, in daily activities—all of these depend on individuals inferring meaning creatively out of processes in the world larger than the person. This very publication, in its commitment to intellectual pluralism, treats sense-making as a human skill, rather than as a product manufactured in volume by an internet algorithm, or an ideology, or a click-bait sweatshop. What are the principles governing such skills, such values, such groping around in the dark?
The standout sequence in Ex Machina arrives in the second half of the movie. Caleb is disturbed at finding that Nathan had ripped up a picture Ava had drawn. Looking for him in the bowels of the complex, he meets Kyoko, who is staring at a Jackson Pollock painting in a large room, but she doesn’t speak. Nathan soon enters behind them (beer in hand), turns on the 1983 disco track, “Get Down Saturday Night,” by Oliver Cheatham, and suggests that Caleb dance with her (“I told you, you’re wasting your time talking to her—however, you would not be wasting your time if you were dancing with her.”) The lights in the room dim to a red glow, with blue-and-white lights strobing behind the decorative lattice-work on the walls. Caleb asks Nathan about the picture, but he ignores the question and starts dancing with Kyoko, while Caleb, horrified, watches on. After a minute or so, the sequence cuts to Caleb and Nathan in the hallway, leaving the disco room after Nathan appears to have blacked out. Throughout the dance sequence the camera cuts relatively quickly between Caleb and the dancers at different angles, and slowly pans in or pans out in each shot. There’s a homage here to Saturday Night Fever, in the palette and the dance moves, but in its terseness and in Caleb’s reactions, the reference feels intentionally absurd and disquieting rather than satirical.
This disco sequence works in the plot (Caleb becomes more motivated to help Ava escape because of Nathan’s erratic response) and is revealing of character (emphasizing, for instance, Nathan’s physicality and his macho, though rather constipated, sexuality), but these details are not why the scene has a such a heightened cinematic impact against any number of other scenes in the movie. The scene’s power comes from certain of its parts—the dancing, the lighting, the music, the camera work, the Jackson Pollock painting in the background, Nathan and Kyoko’s sexual appeal, the incongruity of Caleb’s alarm with the ostensible joyous physicality of the dance, etc.—fitting together in a latent unity, but on the basis of no obvious metonymic or even metaphoric connection. The cinematic glue, so to speak, that makes this sequence an outstanding whole lies in something felt, through these parts, not connected referentially but unconscious, and shaped into an audio-visual spectacle. Oscar Isaac, the actor who plays Nathan, called the sequence a “disco non-sequitur” and I think this epithet gets at the crux of its cinematic unity.
Popular culture in the U.S. has been dominated by the myth of artificial intelligence for at least 50 years (Ava could easily be interpreted as a femme fatale version of HAL 9000 from 2001), and it infects not just entertainment, but what is popular broadly in the economy and in intellectual life. This myth, in its various retellings, has something to teach us about how the automated tools produced in the industrial revolution have impacted our lives. But it has nothing to say about the appeal of Ex Machina’s disco non-sequitur. A promising theory of artificial intelligence, one that might actually allow these systems to behave more like humans, should have an explanation of disco non-sequitur as a fact. Larson shows us, definitively, that nothing like this exists in the contemporary intellectual landscape, to the detriment of our wider culture. It is all the more important to ask, therefore: what would such a theory look like?