Operations Practice An executive primer on artificial general intelligence While human-like artificial general intelligence may not be imminent, substantial advances may be possible in the coming years. Executives can prepare by recognizing the early signs of progress. by Federico Berruti, Pieter Nel, and Rob Whiteman © Getty Images April 2020 Headlines sounding the alarms that artificial University (NYU), is much more direct: “It’s hard intelligence (AI) will lead humanity to a dystopian to explain to non-specialists that AGI is not a future seem to be everywhere. Prominent thought ‘thing’, and that most venues that have AGI in their leaders, from Silicon Valley figures to legendary name deal in highly speculative and theoretical scientists, have warned that should AI evolve into issues...” artificial general intelligence (AGI)—AI that is as capable of learning intellectual tasks as humans Still, many academics and researchers maintain are—civilization will be under serious threat. that there is at least a chance that human-level artificial intelligence could be achieved in the Few seeing these warnings, stories, and images next decade. Richard Sutton, professor of could be blamed for believing that the arrival of AGI computer science of the University of Alberta, is imminent. Little surprise, then, that so many media stated in a 2017 talk: “Understanding human- stories and business presentations about machine level AI will be a profound scientific achievement learning are accompanied by unsettling illustrations (and economic boon) and may well happen by featuring humanoid robots. 2030 (25% chance), or by 2040 (50% chance)— or never (10% chance).” Many of the most respected researchers and academics see things differently, however. They What should executives take away from this argue that we are decades away from realizing AGI, debate? Even a small probability of achieving AGI and some even predict that we won’t see AGI in in the next decade justifies paying attention to this century. With so much uncertainty, why should developments in the field, given the potentially executives care about AGI today? The answer dramatic inflection point that AGI could bring is that, while the timing of AGI is uncertain, the about in society. As LeCun explains: “There is disruptive effects it could have on society cannot be a thin domain of research that, while having understated. ambitious goals of making progress towards human-level intelligence, is also sufficiently Much has already been written about the likely grounded in science and engineering impact of AI and the importance of carefully methodologies to bring real progress in managing the transition to a more automated world. technology. That’s the sweet spot.” The purpose of this article is to provide an AGI primer to help executives understand the path to machines For business leaders, it is critical to identify achieving human-level intelligence, indicators by those researchers who operate in this sweet which to measure progress, and actions the reader spot. In this executive’s guide to AGI, we aim to can take to begin preparations today. help readers make that assessment by reviewing the history of the field (see sidebar, “A brief history of AI”), the problems that must be solved How imminent is AGI? before researchers can claim they are close to In predicting that AGI won’t arrive until the year developing human-level artificial intelligence, 2300, Rodney Brooks, an MIT roboticist and and what executives should do given these co-founder of iRobot, doesn’t mince words: “It is insights. a fraught time understanding the true promise and dangers of AI. Most of what we read in the headlines… is, I believe, completely off the mark.” What capabilities would turn AI into AGI? Brooks is far from being a lone voice of dissent. To understand the complexity of achieving true Leading AI researchers such as Geoffrey Hinton and human-level intelligence, it is worthwhile to look Demis Hassabis have stated that AGI is nowhere at some the capabilities that true AGI will need to close to reality. In responding to one of Brooks’ master. posts, Yann LeCun, a professor at the Courant Institute of Mathematical Sciences at New York Sensory perception. Whereas deep learning has 2 An executive primer on artificial general intelligence Sidebar A brief history of AI The term “artificial intelligence” was around computer vision, natural lan- knowledge of the world through rela- coined by John McCarthy in the research guage understanding, and neural nets tionships between these symbols. One proposal for a 1956 workshop at Dart- are, in many cases, several decades old. familiar example is the knowledge that mouth that would kick off humanity’s a German shepherd is a dog, which is a efforts on this topic. The AI topics that Within these AI research communities, mammal; all mammals are warm-blood- McCarthy outlined in the introduction there has been substantial success ed; therefore, a German shepherd included how to get a computer to use and advancement in what the field now should be warm-blooded. human language; how to arrange “neuron terms “narrow AI” applications. Narrow nets” (which had been invented in 1943) AI is the application of AI techniques to a The limitation is that humans still provide so that they can form concepts; how a specific and well-defined problem, such the ground truth by encoding our knowl- machine can improve itself (that is, learn as chatbots that help customers resolve edge of the world, rather than allowing or evolve); how a machine could form ab- issues with their phone bills, algorithms an AI system to observe and encode stractions by using its sensors to observe that spot fraud in credit-card transac- these relationships itself. Symbolic AI the world; and how to make computers tions, and natural-language-processing was the dominant paradigm of AI re- think creatively. In essence, McCarthy engines that quickly process thousands search from the mid-1950s until the late was describing in 1956 what we now of legal documents. Applying narrow AI 1980s. One ongoing effort to provide a call AGI. solutions in use cases across industries solution to common-sense reasoning can generate tremendous economic through symbolic AI is the Cyc Project, McCarthy was certainly not the first to benefits. Our colleagues’ research launched in 1984 to collect knowledge talk about machines and “intelligence.” shows that the potential value of ap- represented declaratively in the form Alan Turing’s 1950 paper on “Comput- plying this sort of deep learning could of logical assertions, of which it had ing machinery and intelligence” intro- range from $3.5 trillion to $5.8 trillion collected 25 million by 2017. duced the “imitation game,” a test of a annually. machine’s ability to exhibit intelligent Neural networks (1954, 1969, 1986, behavior, and which became known as To differentiate themselves from re- 2012). Recent advances in speech the “Turing test.” Turing optimistically searchers solving narrow AI problems, recognition (now widely used in devices estimated that, in the year 2000, a com- a few research teams have claimed an such as smart speakers), and in com- puter with 128Mb of memory would have almost proprietary interest in producing puter vision (such as limited self-driving a 70 percent chance of fooling a person. human-level intelligence (or more) under abilities in cars) are all due to deep In his earlier 1948 paper on “Intelligent the name “artificial general intelligence.” neural networks. The artificial neuron as machinery,” he describes what we today Some have adopted the term “super- a computational model of the “nerve net” call computers, as well as a machine that intelligence” to describe AGI systems of the brain was proposed as early as fully imitates a person. He points out that by themselves could rapidly design 1943.¹ In the decades since then, it has that our ability to build adequate sensors even more capable systems, with those been through multiple highs and lows in and actuators might not be sufficient for systems further evolving to develop ca- its popularity as a tool for AI. some time and that our efforts are best pabilities that far exceed any possessed invested in the aspect of intelligence that by humans. MIT’s Marvin Minsky and Seymour relates to games and cryptography. The Papert put a damper on this research in clear aspiration, however, has always MIT roboticist Rodney Brooks describes their 1969 book “Perceptrons,” where been to achieve human-level intelligence. the four previous attempts at AGI, along they mathematically demonstrated that with their approximate start dates, in neural networks could only perform very The early work, like Turing suggested, discussions that provide important con- basic tasks. They also discussed the dif- revolved around subject areas that do text for understanding their progress. ficulty of training multi-layer networks. not require too much sensing and action, In 1986, however, Geoffrey Hinton and such as those of games and language Symbolic AI (1956). The key concept is some colleagues solved this problem translation. Research communities the use of symbols and the encoding of with the publication of the back-prop- 1 Warren McCulloch and Walter Pitts, “A logical calculus of the ideas immanent in nervous activity,” Bulletin of Mathematical Biophysics, volume 5, 1943. An executive primer on artificial general intelligence 3 Sidebar cont. agation algorithm. In the 1990’s, work vision to identify and navigate through Behavior-based robotics (1985). by Yann LeCun made major advances in their environments or to understand Insects navigate well in the real world neural networks’ use in computer vision, the geometry of objects and maneuver with very few neurons. This observation and Jürgen Schmidhuber similarly them, such as moving around blocks of inspired researchers to investigate how advanced the application of recurrent various shapes and colors.