by Angus McIntyre
Angus McIntyre discusses the past, present, and future of artificial intelligence in science fiction. Read his latest story,”Boojum,” in our [September/October issue, on sale now!]
It’s probably true to say that many of the turning points in human history passed unnoticed by those who lived through them. Nobody woke up one morning and was startled to find that the Roman Empire had fallen. The Fall wasn’t a discrete event, but a long decline that spanned decades or centuries. And while intense events like world wars are difficult to overlook, if you’re in the middle of one it’s hard to say exactly which direction history might be taking.
But I would cautiously suggest that we might now be at some kind of inflection point, a time when we can see a “before” and “after.” We know that the “after” is going to be very different from the “before,” even if we don’t yet know precisely what shape it will take. And it’s because of the thing we’re calling “AI.”
A quick disclaimer: “AI,” as the term is being used now, is a terrible term. It’s uninformative and misleading, and mostly marketing hype. A less loaded and more accurate label might be “machine learning” (ML), but it may be too late to substitute one term for the other in everyday use.
My own involvement in the field of artificial intelligence started shortly before what I like to call the Second Great Disappointment. Artificial intelligence research has gone through several cycles, each one starting with the conviction that intelligent machines were just around the corner, and ending with the realization that they very much weren’t. After each Disappointment, the field reset, those involved retracted some of their wilder claims, and the cycle started over.
The First Great Disappointment came in the 1960s or 1970s. The preceding decade had seen the emergence of digital computers, a new kind of machine built to execute not physical but intellectual tasks. The newly-minted computer scientists began to wonder if we might be able to build intelligent machines.
First, of course, they had to define what intelligence was and what it meant for a machine to be intelligent. Sensing they were about to get lost in the philosophical weeds, early researchers in what was then called ‘machine intelligence’ adopted a rule of thumb: an intelligent machine was one that could do tasks that would require intelligence if done by a human. By this, as it happened, they mostly meant planning and reasoning—tasks that many actual humans go out of their way to avoid.
At first, there was optimism. Using the astonishingly powerful computers of the day—room-filling behemoths a few orders of magnitude less powerful than the smartphone in your pocket—the goal seemed eminently achievable. Creating an intelligence was probably just a matter of writing a lot of extremely clever programs and hitting “RUN.”
This turned out not to be the case. The machines were too slow, the tools too weak, the problems just too hard. The First Great Disappointment had arrived.
By the time I went to college to study Linguistics and Artificial Intelligence, the field had reset. We now had faster computers and specialist programming languages like LISP and PROLOG. We had a better understanding of the issues, and a confident conviction that if we could just encode enough human knowledge in formal rulesets, intelligence would follow.
It didn’t. As the Second Great Disappointment loomed, the field split between proponents of “strong AI”—who still clung to dreams of building an artificial general intelligence (AGI)—and “weak AI.” “Weak AI” was about building systems that could display human-like competence at very constrained tasks. Such ‘expert systems’ found practical applications in medicine or industry, while AI-related techniques such as search were quietly subsumed into mainstream computer science.
Artificial intelligence research has gone through several cycles, each one starting with the conviction that intelligent machines were just around the corner, and ending with the realization that they very much weren’t.
Science-fiction writers had mostly echoed AI researchers in portraying intelligence as something that could be programmed. Asimov’s famous Laws of Robotics implied programming: the idea that intelligent entities could be controlled by easy-to-understand rules hard-coded into their brains.
One who didn’t was Roger Zelazny. In his 1976 novella, Home is the Hangman, he described an intelligent machine whose architecture mimicked the human brain. The Hangman wasn’t programmed; it learned.
Zelazny had probably been talking to a few researchers in an emerging field known as connectionism. “Traditional” artificial intelligence focused on symbolic approaches, where facts and concepts were represented by manipulable symbols. Connectionism was sub-symbolic, trying to build a brain by imitating its smallest parts. Zelazny’s summary of the core concepts was so exact that the first time I read about connectionist neural networks I wondered why the concept seemed so familiar.
Neural networks looked more promising than the symbolic techniques I had learned. But neural nets weren’t really delivering either. The hardware was too slow, the models too small, the datasets too limited. There was enough disappointment for everyone.
Today, however, things have changed substantially. Sub-symbolic approaches have come of age. Optimized hardware allows us to execute the trillions of simple calculations needed to train non-trivial statistical models in acceptable time. Giant data sets (not all gathered with the consent of the data owners) provide training data on the scale needed for effective learning. The core algorithms have been optimized and extended. “AI” is everywhere, performing impressive party tricks such as creating images in response to simple descriptions, or engaging in eerily human-like typed conversations. Are we, after all, on the brink of creating a true artificial general intelligence?
Some of the more excitable voices in the field say “Yes.” They also say that there’s a good chance that any AGI we build will wipe out humanity, which isn’t reassuring. At least none of my PROLOG programs ever looked likely to exterminate the species.
These ‘strong AI’ enthusiasts are probably wrong again, at least for now. I doubt we’ll see Clarke’s HAL-9000, Banks’s Culture Minds, or—fortunately—Ellison’s AM any time soon. The reality is that today’s “AI” products are often far more limited and error-prone than their makers would admit. Attempts to tweak some popular systems have actually made them worse at what they do, while ChatGPT and friends have been revealed as glib pathological liars, fundamentally incapable of telling fact from plausible-looking fiction. Another Great Disappointment is coming, sharpened by the gulf between the breathless hype and the less-impressive reality.
We’re also waking up to some real concerns about the ethics of the technology, ranging from the questionable sourcing of the data needed to build the models to the human impact of deploying these untested and imperfect systems. The blind rush to adopt tools that are at once too good and not good enough at what they do will come with real human and societal costs. It’s probably just as well that we’re nowhere close to creating ‘real’ machine intelligence. If we can’t cope with the ethical problems involved in unleashing a bunch of dumb-as-rocks chatbots on our world, we’re in no way ready to deal with the moral and practical dilemmas associated with creating an entirely new class of sentient entities.
But despite the ethical concerns, even though we’re not much closer to the “strong AI” goal of a truly general intelligence, today’s machine learning systems represent a genuine breakthrough and one that is likely to change our world in profound ways.
Over the next few decades, we will see the products of machine learning become ubiquitous. Machine learning is not a fast track to the goals of strong AI; rather, it is a triumph of weak AI. Instead of creating general intelligences, it promises to open the door to a world of specialist intelligences, pseudo-intelligent systems that exhibit human-or-better levels of performance in highly constrained domains.
We can, in essence, take almost any field of expertise—the more specialized the better—and encapsulate it as a black box that accepts inputs and gives back appropriate outputs. We can even teach our black box to solve problems we don’t know how to solve ourselves; all we need is a large enough set of input data and an idea of what we want to get out of it. The learning algorithms will figure out how to get from one to the other.
There are, of course, an immense number of caveats. As ever, the devil is in the details and assembling good training data is one of the most devil-ridden parts of the whole enterprise: garbage in still means garbage out. But while today’s crop of chatbots and image generators may prove to be short-lived curiosities, the technologies that spawned them will not.
The genies are out of the bottle and looking for work. As the cost of developing new models falls, and the hardware needed to support them becomes more widely available, adding these artificial specialists to products will become a default choice. It won’t always be the right choice, but marketing departments have rarely been deterred by questions of what’s useful or appropriate.
At first, they will mostly be embedded in other pieces of software, a process that has already begun. Apps running on smartphones or computers will be augmented with “AI” features. Specialist applications based on machine learning models will find niches in business, science and medicine, as well as other domains. But sooner rather than later, ML models will make the jump to hardware.
Many are likely to be invisible to the end user, hidden enhancements to make a complex device work more efficiently or reliably. But others may be interactive. Enhanced devices might share diagnostic information or ask for instructions. Smart tools could assist in their own troubleshooting or provide on-the-job training.
Speech will likely be the interface of choice (although manufacturers will need to provide an alternative for deaf users). Chips providing basic speech recognition and the ability to learn conversational models for a simple domain will be off-the-shelf components in a few years. Instead of trying to control your smart device through a non-standard visual interface full of cryptic icons, you’ll be able to talk to it through your smartphone or home network (larger devices may also support embedded speakers and microphones).
What this means is that before long we may find ourselves living in a world populated by things that we can talk to much as we talk to other people. Animists believe there are spirits in every tree, hill or waterfall. Once the building blocks become commoditized, our world will be filled with similarly inhabited objects. You will be able to have a conversation with a traffic light, a washing machine or a lathe, learn how it is feeling and tell it what you need it to do. Buildings and places can be endowed with their own artificial genius loci. And they won’t only be talking to us; networking will allow them to talk to each other, with individual devices forming nodes in a larger entity whose behavior may in turn be determined by other, still more specialized models.
Even for people who might like to live in a world of ‘ensouled’ objects, this will not be entirely idyllic. Today’s ML models are unpredictable. Like the expert systems of the 1980s they are ‘brittle’: anything that takes them outside their core area of expertise can lead to dramatic failures. Attempts to manage their behavior by adding “guardrails”—21st-century equivalents of Asimov’s Laws — have had mixed success at best. Moreover, the cheapest embedded models are unlikely to receive updates, so any flaws in the original model will persist for the lifetime of the device. A world of smart objects will also be a world full of devices with baked-in quirks, or which can be subverted in surprising or even dangerous ways.
Machine learning will change our world in other ways too. Regrettably, surveillance and the military are likely to be early beneficiaries. And if experience is any guide, a machine that can strike a target with superhuman accuracy may not be quite as good at distinguishing combatants from civilians.
Science-fiction has most often depicted the products of strong AI: large, human-like general intelligences that function like people (or minor gods). Portrayals of a world full of ubiquitous micro-intelligences, tiny, embedded specialists whose superficial articulacy hides their constrained scope, are harder to find. But I believe that’s the model for the world we will soon be living in, and it’s fertile ground for any writer who wants to tell that story. It’s a world full of things that talk, things with their own quirks and their own agendas, things that can be conjured and coerced. At times, it may seem more like writing fantasy than science-fiction.
But anyone who sets out to describe that world will need to write fast. It’s already on its way, and it will be here sooner than we think.
The Astounding Analog Companion 