It’s our pleasure to welcome you to the University of Toronto for our fifth annual conference, Machine Learning and the Market for Intelligence.
Last year, our conference theme was complements – the ingredients necessary, other than models, to make commercial-grade AI work: data, hardware, redesigned workflows, regulation, human judgment, infrastructure, workforce training, etc. The year before, our theme was time – a shift in the narrative from “if” AI would transform key business products, services, and processes to “when” – and the implications for investment strategies. This year, our theme is power.
Twenty years ago, in the market for information curation, there were many small centers of power. We called them libraries. Every town had one. Some had more than one! Although a few were especially grand, like the Library of Congress in Washington, the Bodleian Library at the University of Oxford, and the Vatican Library in Rome, from an organizational and geographic point of view, power conferred by the curation of information was fragmented and broadly distributed across thousands of libraries.
As recently as 1998, when two Stanford students, Larry Page and Sergey Brin, were incorporating their new venture among a gaggle of other internet search-related companies, almost nobody predicted that twenty years later the world would drastically change and that power in the market for information curation would shift away from the many libraries scattered across the globe and consolidate so spectacularly into the hands of such a small number of corporations, such as Alphabet (Google), Microsoft (Bing), and Baidu.
Furthermore, twenty years ago, the high priests in the information curation world were librarians, often with advanced degrees in library science. At that time, there was seemingly no discussion of a dramatic shift in power from library science-related to data science-related occupations.
Moreover, effectively no public policy discussions at that time anticipated a shift in the power of nations with respect to information curation that today is the topic of many antitrust and national security debates as a result of power that moved from many libraries distributed across all countries to search companies primarily domiciled in just two nations: the United States and China.
Finally, despite the concentration of power with respect to the curation of information, power conferred by access to information has never been so broadly distributed. In particular, poorer neighborhoods and countries that were previously disadvantaged due to their limited access to well-funded libraries now benefit significantly from this change in regime.
Several important factors led to the shift in power with respect to the curation of information. We focus here on one of them that continues to play a key role to this day: artificial intelligence (AI).
While many applications of AI enhance productivity by improving upon existing predictive analytics, such as in banking with applications in areas such as fraud detection, anti-money laundering, know-your-customer, credit scoring, and sanction screening, some applications have a more profound effect: they shift power.
Certain skills become more important, shifting power from some occupations to others (e.g., from library science-related to data science-related occupations). Certain industry-level characteristics become more salient, shifting power from some organizations to others (e.g., from libraries to software companies). Finally, certain country-level characteristics become more critical, shifting power from some nations to others (e.g., from all countries to just a few, the US and China in the case of search).
Another example is private transportation. The power in this industry used to be fragmented and distributed – many taxi companies across every country. Drivers invested in skills, especially knowledge of the city’s roads and traffic patterns. In the City of London, for example, drivers of black cabs are required to learn The Knowledge, which requires three years of schooling and creates a barrier to entry into the profession. Now, navigational AIs enable newcomers to land at Heathrow airport, rent a car, and navigate the city with virtually the same level of sophistication. This, along with dispatching AIs that predict the supply and demand for rides, enabled a dramatic shift in power from thousands of taxi companies around the world to a small handful of powerful software firms, such as Uber, Lyft, DiDi, Ola, and Hailo, that control a staggering percentage of rides around the planet.
The more users these systems attract, the more data they collect. The more data they collect, the better predictions they can deliver. And of course, the better predictions they can deliver, the more users they attract. As the flywheel gains momentum, one or a few companies pull ahead of the rest, and the competition eventually falls behind and cannot catch up. Machine intelligence leads to winner-take-most outcomes and hence power.
We are just starting to see investors and large corporations begin to move their pieces around the chessboard in anticipation of these power shifts. Take, for example, medical diagnosis – a prediction problem. Progress is impressive in this area: AIs better able than doctors to classify conditions in pathology slides, AIs better able than radiologists to identify conditions in medical images, and AIs better able than mental health experts to predict suicides.
So far, most of these AIs are early stage – reported in scientific journals but not yet deployed at scale in clinical settings. But many see the writing on the wall. Once one or two diagnostic AIs develop a performance lead over the others, who will want to use the second-best diagnostic AI when the best is no more expensive? The marginal cost of a high-quality prediction is no greater than that for a mediocre prediction. As more healthcare professionals gravitate towards using the leading diagnostic AI, and as long as they provide the system with feedback on its predictions, then the better it will get. It’s not inconceivable that in less than 10 years, power associated with diagnostics will have shifted from hundreds of thousands of doctors’ offices around the world to a small group of AI software diagnostic firms, likely domiciled in just one or two countries. Power in medical diagnosis could consolidate and dramatically impact the structure and dynamics of the entire healthcare system.
When AIs simply improve upon existing predictive analytics or incrementally improve existing processes, power structures do not change and the strategic implications are limited. In these cases, success depends primarily upon well-managed implementation. The objective is enhanced operational efficiency. Decisions concerning when, where, and how to deploy these AIs require oversight from the COO, but not the CEO. These AIs are tactical, not strategic.
However, when AIs lead to shifts in power, the implications for strategy can be significant. Often, the owner of the AI will benefit from the shift in power (e.g., power shifts from many doctors to a small number of diagnostic AI software companies). If the AI is commoditized, however, then the AI owner will not be the beneficiary. Instead, power will shift to those who control the complements (e.g., hardware providers for medical imaging or treatment centres). Unless the complements are commoditized. If both the AI and the complements are commoditized, then the benefits flow through to customers (e.g., patients).
This may sound familiar. It is! The general idea isn’t new. In fact, it’s predicated on decades of research in a field of economics known as Industrial Organization. Furthermore, the application of this field to business strategy was popularized forty years ago by Harvard professor Michael Porter via his Five Forces framework for assessing the profitability of an industry as a function of market structure. The twist here is that market structure is changing as a result of a sudden drop in the cost of prediction due to advances in artificial intelligence.
As in prior years, our focus at this conference is not on the technical details of machine learning – the underlying mathematics, statistics, and algorithms – but rather on the economic and social implications of commercial-grade AI. Hence, our decision to focus on power. Today, we’ll explore state-of-the-art machine intelligence in several domains against the backdrop of power. How are recent advances in machine intelligence poised to impact jobs, markets, and nations via strategies crafted for deploying capital, allocating labour, competing for customers, and enhancing society?
Reflecting on the rate and direction of progress in AI since we founded this conference in 2015, it seems increasingly likely that these will be the questions of our age.
Founder, Creative Destruction Lab
Professor, University of Toronto
These ideas are based on a forthcoming essay coauthored with Joshua Gans and Avi Goldfarb.