|
Prologue Randy P. Sparkman September, 2023 All Rights Reserved I’ll begin with the end. Our machines are going to start to behave like humans. To be precise, they are going to take advantage of advances in artificial intelligence in ways that let them appear to interpret and reason. And, they are going to do that in a conversation - with us. As they do, they are going to become even more useful. They are going to amplify the best and worst of human behavior. And, they will continue to reshape society in profound ways. Even though they will mimic and reflect our manner and do things we cannot–as a hammer drives a nail–they are not going to become human in the ways we understand and that defy definition. The media, government, and technology industry, cheered on by true believers and doomsayers alike, are in a dither. They are convinced, and want to convince us, that a recent consumer product, ChatGPT, is proof artificial intelligence has reached a tipping point that will create utility and disruption equal to the printing press and electricity. Further, because this technology exhibits a sort of human behavior, some believe it's the beginning of the much-predicted, end-of-days role reversal—where technology becomes the master and us, its servant, or, perhaps even its victim. We knew all along that machines were going to don our cloak. We could not help but make them in our image. World War II codebreaker and computer science pioneer, Dr. Alan Turing, explained this in 1950. We heard it from the four horsemen of science fiction: Isaac Asimov, Arthur C. Clark, Ray Bradbury, and William Gibson. We saw it at the movies. We’ve heard it from the technology industry and our crazy uncle. But, none of that stopped us, and the rest of the planet, from being shocked when ChatGPT began to speak. Once presented with that magic, we were quick to apply all those Olympian predictions to the first artificial intelligence product to get the Madonna treatment. Yet, does the sudden appearance of ChatGPT really herald the arrival of the self-aware, self-interested machine? Or, will it, instead, prove to be something less existential, perhaps more steam engine than iceberg, a watershed advance in our digital future equal to the Internet and the smartphone? Or, will it turn out to be an important, but ultimately less thunderous, platform on which Microsoft and Google build buttons and pull-downs and more mountains of cash? Breakout ChatGPT is a form of “generative” artificial intelligence. These systems “generate” various forms of digital output–words, images, video, computer programs–in response to text input from a user. They create new, unique content by comparing user input with patterns and structures from digital information on which they are trained. The “chat” interface allows a user to explore these patterns in conversation, with natural language as both input and output. The result is a rich, two-way exchange of information between user and computer that creates reliable utility and, sometimes, sparks surprising and mutual creativity between human and machine. ChatGPT was developed by a San Francisco startup, OpenAI. They leveraged recent, unprecedented advances in artificial intelligence software pioneered by a team at Google, who made their advances available to AI researchers and developers at large. The OpenAI team used the AI software technique Google refined to create a large language model–that is, a mathematical representation of text patterns stored on a computer. The model, which OpenAI calls “Generative Pre-Trained Transformer” (GPT), is trained on vast amounts of data stored on the Internet and elsewhere. OpenAI created incremental versions of GPT with increasing power over the past five years, culminating in the current version, GPT-4. This latest version is by far the most capable and has become, so far, the standard by which competitor large language models are measured. Through their development process, the OpenAI team learned that as they increased the amount of data on which this model was trained and, at the same time, increased the power of the computers on which it executed, the model became exponentially more capable. It was this “scaling” of computation and data combined with the newly advanced AI software that caused the system to exhibit unexpected capability. The engineers at OpenAI were, in fact, surprised at the extent to which the model began to exhibit human-like reason and insight. And, they remain unable to completely explain why the system can do the things it does. The original AI team at Google, and other AI researchers who began to work with it, realized that they were seeing something quite new and significant, with implications to match. There was a measure of consensus in the AI community that user-facing product development should be slow and deliberate to ensure there was maximum understanding of what these systems could do and the impact of their release into the world. Their caution was prudent, but someone was going to put this technology into a user-facing consumer product sooner or later. OpenAI said that they released a product to allow the planet to incrementally absorb both the positive and negative disruption of the technology. In any case, we got a preview, and they claimed first-mover advantage in the process. Language Language is fuel. It sustains our exertions, connections, and understanding. It is also the central expression of human thought, cognition, and intelligence. From the earliest days of artificial intelligence research, developers understood this and worked to build systems capable of creating and understanding human language. They sought two things: to emulate human intelligence and to provide utility in the process. The makers of generative language AI, of which the company OpenAI is a momentary leader, have achieved both. They figured out how to take the vast digital linguistic output that now spans almost every human endeavor and play it back to us. When we write, we express human thought. That thought is most often, now, captured in digital form. As a result, the ocean of data that has been stored on the Internet is effectively a vast digital representation of human cognition and creativity. This collection not only embodies patterns of language–meaning, rules, structure–it also embodies subtle patterns of how we, as humans, arrived at that output–the way we think. It’s these patterns, both structure and cognition, that these new artificial intelligence systems are extracting. The developers of generative language AI haven’t made something equal to, or greater than, human intelligence–so far–but they have found a way to package the output of the real thing in a way that seems to think and promises to disrupt. What Can It Do? Think of a large language model as a well-mannered college student who has memorized all of Wikipedia, which effectively encompasses all the domains of human endeavor. He has been made available to you as an assistant. His primary job is to fetch information, to help you express that information, and to serve as tutor and coach as you request. He is available at any hour. He has infinite patience. He will engage in conversation with you as long as you want. He will respond to you in any form you request: short, long; summarized, expanded; narrative or bullets; compared or contrasted; simple or complex. He will brainstorm with you on any topic. He will originate ideas and suggestions, in multiples. He will critique and edit your ideas. He will be quirky. He will always provide a response of some sort. Sometimes, but not routinely or predictably, his facts will be wrong, or, he may appear to have a flash of insight. His response is unique to the moment and specific to your need; ask the same question a second time and the answer will be the same, but somehow different. But, this is not a college student, it’s a machine. The technology that enables ChatGPT doesn't "understand" information in the way of humans. Rather, it has ingested and analyzed enough examples of human language to be able to predict, and assemble into sentences, words as a best guess at an appropriate response to a query. These systems are an expansive reference book, encyclopedia, manual, textbook, and technical document. They can be trained on any kind of information across any domain. But, significantly, they only know facts on which they are trained. They will help you share information and express ideas. You can take the facts the system has provided and craft your own narrative, or you can ask the machine to craft one. You can provide your own words, facts, ideas, and opinions and the system will partner with you to create that story. It will iterate with you as you create. At each step, this iterative process will sometimes spark deeper insight or an unexpected connection. Language AI systems can also be deployed to use these language skills as a teacher, coach, or companion, and, sooner than later, as a personal digital command center. All these capabilities have one thing in common: they augment and amplify your activity–human activity. This is a tool, an extension of our ability. They don’t have meaning or value without human engagement and effort. That additive value to your own unique baseline of ability across any domain is the headline and the thunder of these systems, far more than any sensational news about a mistake or their ability to write a haiku about a spaceship. Leap Frog One could argue that this notion of human augmentation is not new. All the digital technology of the past fifty years: e-mail, word processors, Internet browsers, smartphones, and the rest have been our helpers. In his influential 1995 book “Being Digital”, Nicholas Negroponte, then head of the MIT Media Lab, famously said, “Anything that can be digitized will be.” He was correct, that happened, and it has reshaped how we work, learn, and live. What we see now with generative language AI is a step-advance in the digitization that Nicholas Negroponte predicted. ChatGPT would not exist without that vast repository of digital information–human capital in the form of expressed intelligence–on which it relies. We’ve seen the capability and use of systems that generally fall into the category of “artificial intelligence”–of which generative language AI is a category–increase at an exponential rate over the last decade. These include virtual assistants, prediction and recommendation systems, robotics, and healthcare diagnostics. Language processing has been the vanguard of those advances. Up to now, we experience the result, but we don’t see behind the curtain. We’ve all had the “big brother is watching” feeling with our smartphones and online shopping. We’re not certain, but it feels like someone is scooping up our written and spoken words and playing them back to us in advertisements and recommendations. And, in fact, they are, with the help of language AI. Amazon, Google, Facebook and their little brothers have so far kept these AI tools under the hood. We reach the destination, but our hands aren’t on the wheel. OpenAI raised that hood with ChatGPT. They presented to us, as everyday users, an incarnation of the current state of language processing tools with a powerful-in-its-simplicity chatbot user interface. Not only did they show us the technology, but they also let us drive. Frontier We are in the earliest days of consumer use of AI-powered digital information. OpenAI dropped onto our doorstep an unrefined version of what’s to come. We certainly see enough to begin to appreciate the leap, and we can begin to understand what it can do. Yet, we are nowhere near a place to fully appreciate the shape these tools will take or their ultimate impact, nor can we be. That will be a function of us; our use will shape the products. When presented with something new, our reflex is to map to what we already understand. We will certainly do that with language AI. We will compare its function to the function of the digital tools that emerged over the last generation. We will assume and expect language AI to be an improvement in what we are already familiar with. There is certainly overlap. For one thing, language AI would not exist without that foundation. Computer networks, office tools, social networks, PCs, data storage, cloud computing, digital literacy, and the Internet itself, had to be in place for language AI to emerge. We will expect language AI to perform information retrieval, that most basic computing function. It will, at some point, help us create, store and share documents. But, in many ways, language AI is a left turn in what has been a straight-line progression of digital capabilities. The Internet was about frictionless digital information, sharing that information and using it to connect with others. Much was written about the idea of disintermediation. With friction reduced, we were suddenly able to bypass many gates. Before the Internet, musicians needed record labels to sell their music. Now, musicians share their music directly with listeners, the role of the middleman much reduced. Language AI is something much different with the potential to reorient that transaction. In fact, it will reintroduce a kind of mediation. Now, it will extract and, in some cases, create digital information on our behalf. We will be in the loop. We will and must, in fact, have the final say, but, suddenly, in something of a post-Internet way, we have an Assistant working between us and the source of information. The shape that takes…? Who knows? Comes Now the Golem? ChatGPT was so startling, and our pop culture preconceptions about self-aware and self-interested machines so ingrained, that discussions about how we will use and manage these language systems were immediately drowned out with questions of “Will this thing eat us?” ChatGPT can mimic human behavior just enough to conjure the notion of a sentient and conscious machine, however you choose to define those words. Add to that, developers themselves cannot completely explain what enabled this leap in function and why, sometimes, it seems brilliant, and other times, confused. But, discussions of other significant but less existential concerns did finally begin to emerge. The cautionary tale of the downside of social media is not lost on any of us. As an amplifier of human agency, generative language AI has the potential to expand opportunities for human mischief. Issues that bedevil our current digital tools like intentional misinformation, bias, breaches of privacy, fraud, and disruption of jobs will almost certainly be made worse by tools that, by definition, enhance our exertions, even the bad ones. Not So Fast What we see now with ChatGPT is more preview than product, more prototype than premier. It is a very long way from becoming a conscious being with its own agenda. At this point, it’s more parrot than Yoda. The proclamations of evangelists and the hand-wringing of doomsayers aside, we will process the promise of this technology as we have others, in uneven ways, with false starts and flashes of innovation that finally give way to usefulness. If we fear shadows on the wall or fail to be honest with ourselves about the ways we, not machines, behave, then we risk diminishing the opportunity and utility of an innovation with the potential to improve our day jobs and perhaps our lives, and, most importantly, the lives of our children. The art is, as always, to separate signal from noise, which may, in fact, prove to be the real and ultimate magic that language AI has to offer. |