What to Learn from US Govt Strategy on AI
[This post was co-authored with Ajay Agrawal and Avi Goldfarb. A shorter version was published in HBR Online on 21st December 2016. Also, the post does not review a new White House paper on AI and its impact released on 20th December 2016, that cites some posts on this blog.]
On October 12, 2016, President Obama's Executive Office published two reports that received less media attention than they might have otherwise because the United States was gripped by the final weeks of a presidential campaign race. In these two reports, the administration laid out its plans for the future of artificial intelligence (AI). Depending on one's view of AI's potential impact, the actions resulting from these reports may be more influential on the long arc of history than the outcome of that election. The combined reports include eighty-eight pages and twenty-five recommendations. What does all this mean for other countries? We summarize the seven most important insights for governments of other countries, such as Canada, that collaborate, trade, and compete with the US.
#1 – The US government's approach to AI reflects a sense of urgency.
The two reports were prepared with remarkable speed, especially relative to the traditional pace of government. On May 3, 2016, the White House announced the formation of a new National Science and Technology Council (NSTC) subcommittee, "Machine Learning and Artificial Intelligence," along with a series of public workshops. Six weeks later, on June 15, 2016, that subcommittee directed another subcommittee to draft a separate AI plan focused on R&D. Four months later, the two subcommittees each published a substantive report, both of which articulate a series of recommendations for action. Thus, in only six months, the US government set a priority, enlisted key individuals, and produced a comprehensive national plan for AI.
#2 – They're not treating AI as a science project; it's commercially important.
The plans call for a significant increase in both the human and financial capital allocated to AI. The wide-ranging importance of AI is reflected in the people assigned to the NSCT Subcommittee on Machine Learning and Artificial Intelligence, which includes representatives from the Departments of Commerce, Treasury, Transportation, Energy, Education, Justice, Health and Human Services, Labor, State, as well as the Department of Defense, Department of Homeland Security, Central Intelligence Agency, National Security Agency, and the National Security Council. Furthermore, in a separate report, the Chair of President Obama's Council of Economic Advisers, Jason Furman, wrote: "The biggest worry I have about AI is that we will not have enough of it, and that we need to do more…" Echoing this point, President Obama subsequently remarked: "The analogy that we still use when it comes to a great technology achievement, even 50 years later, is a moonshot. And somebody reminded me that the space program was half a percent of GDP. That doesn't sound like a lot, but in today's dollars that would be $80 billion that we would be spending annually … on AI."
#3 – China is a leader, not a follower or a copycat.
Perhaps the most perplexing part of the report is the stark contrast between a dramatic illustration and the accompanying blasé text. Figures 1 and 2 (copied below) plot the number of scientific and technical publications mentioning "deep learning" or "deep neural network" by nation over the nine-year period 2007-2015. Although slightly different approaches are used to count publications, both figures illustrate the same striking finding. Over the past four years, the US and China grew their research output in AI significantly faster than other countries, with the US initially emerging as the worldwide leader. However, over the last two years, China surpassed the US by these measures of research output. Given the importance of AI reflected in every other aspect of this report, it's surprising that the rise of China and implications for competition in research and human capital receives only one sentence in the text: "The trends also reveal the increasingly global nature of research, with the United States no longer leading the world in publication numbers, or even publications receiving at least one citation." That's all the report says. China is never explicitly mentioned even once in the text of either report.
Given the likely increasing returns in this area such that small leads are likely to grow into larger leads due to network effects (better AIs attract more users, which in turn generate more data and further improve the AIs), localized learning spillovers, and virtuous circles attracting talent (stars attract other stars), it's surprising that both reports are silent on the topic of competition with China. Perhaps the authors of the report believe the data reported in Figures 1 and 2 are not good measures of meaningful research. Alternatively, perhaps the real policy response (an aggressive international recruiting campaign for top AI research talent) is not meant for public consumption. The founder of Chinese company Baidu, Robin Li, expresses his view on the importance of this technology for China: "when the age of AI arrives, the [internet of things] will become a big market and completely change manufacturing. I think that in the future all manufacturing will be a part of the AI industry… China is a manufacturing giant, and I think we need to really pay attention to AI tech development…" Every country with AI research expertise that wishes to participate as producers, not only consumers, on the frontier of this technology and the associated economic and social opportunities should prepare for intensifying labor market competition, especially from the US and China.
Figures 1 & 2 from The National AI R&D Strategic Plan (p. 13)
#4 – There is no clear vision regarding where to focus research funding.
The R&D report notes that the private sector is investing heavily in AI and at an increasing pace, such that it is important for the government to focus their resources on the types of AI research that the private sector will be less likely to support: "…this plan assumes that some important areas of research are unlikely to receive sufficient investment by industry as they are subject to the typical underinvestment problem surrounding public goods." (p. 6) However, although the report provides a long list of research topics, it does not delineate which are commercial versus requiring government support. Perhaps that is because it is so difficult to predict which areas of research are unlikely to have reasonably immediate commercial value.
At present, AI is a field where the line between basic and applied research is so blurred and the future so poorly understood that even the most advanced governments have little insight other than the generic theme to focus on "fundamental and long-term" R&D. This is consistent with the observation that several of the most advanced companies in the field uncharacteristically appointed academics to lead initiatives critical to their business rather than recruit full-time employees to do the job (e.g., Yann LeCun – Director of AI Research at Facebook and Professor at NYU; Andrew Ng – Chief Scientist at Baidu and Professor at Stanford; Ruslan Salakhutdinov – Director of AI Research at Apple and Professor at CMU). This lack of clarity presents an opportunity for other countries. It is precisely how a small country like Canada was able to compete with the US – a country ten times its size – to become a leader in the development of two important subfields of machine learning: deep learning and reinforcement learning. Nations that back domestic star scholars or junior researchers with a promising trajectory who are focused on unique approaches to fundamental problems may develop expertise in areas that are critical to the advancement of the field and thus earn a competitive position in the affected markets and industries.
#5 – The US government has a good roadmap for AI infrastructure.
While neither report offers clear insight into how to allocate resources to research, the reports do offer a series of compelling approaches to other forms of support for AI that are likely to be underserved by private-sector investment. The four most salient are:
1) Regulation: The bottleneck to deployment of AI-enabled, society-enhancing products and services in a number of domains is shifting from technology to regulation, such as those associated with the use of autonomous vehicles and unmanned aircraft systems. The report notes: "where regulatory responses to the addition of AI threaten to increase the cost of compliance, or slow the development or adoption of beneficial innovations, policymakers should consider how those responses could be adjusted…" Well-designed regulations will influence the rate and direction of innovation by creating incentives for the private sector to invest along trajectories that will most benefit society. For example, regulations regarding transparency (provision of an explanation for an AI-generated determination, e.g., "Why is the AI recommending a treatment that is different from what the human doctor recommends? What is its reason?"), efficacy, and fairness, based on evidence-based verification and validation, will create incentives for the private sector to invest in research to address these issues (e.g., the provision of transparency), which will make AI more governable.
Regulation offers another area for regional differentiation and competition. For example, jurisdictions that are first movers in regulations that are friendly towards self-driving cars may attract significant investment, not only from car companies, but also from complementary industries (e.g., intelligent road infrastructure, delivery services, autonomous fueling services, disabled person services, etc.), even though that region is not a leader in AI research. The many regulations that will be challenged by AI-enabled technologies in industries as diverse as medicine, finance, transportation, safety, and justice will create significant opportunities for non-market competition. Countries do not need to be leaders in AI research to compete on this dimension.
2) Education: The reports recommend investing in developing AI talent at all levels and also investing in a data-literate citizenry that will enhance the value of all humans in an economy that employs an increasing level of machine intelligence. The returns to investment in education will vary across countries, depending on the extent to which their labor force participates in industries that either produce or consume AI-enabled products and services.
3) Public Service: The government can use AI itself to serve the public faster, more effectively, and at lower cost. This includes mundane tasks such as faster bureaucratic processes (e.g., issuing driver's licenses) as well as complex applications such as cybersecurity (lower cost, more agile) and weapon systems (safer, more humane). This will be an important area for competition across jurisdictions. National and regional leaders who understand the opportunities made available through the development of AI will incorporate that into their strategies and invest accordingly. Those who invest early may benefit not only from increased productivity but also from learning about how AI-enabled products and services work that may provide additional advantages in subsequent strategy decision-making. Of course, early adopters also risk exposure to the costs of learning and less-tested technologies.
4) Information: The reports call for the US government to "monitor the state of AI in other countries" and also calls on industry to "keep government updated on the general progress of AI in industry, including the likelihood of milestones being reached soon" (p. 41). These recommendations underscore the unpredictable and nonlinear nature of AI research. The committee is concerned that sudden significant increases in AI capability are likely, have non-market implications (security, massive shifts in the distribution of wealth, etc.), and may emerge in other countries or in the private sector. It is in every country's interest to follow this sage recommendation to "monitor the state of AI" as developments could be far-reaching and invoke significant economic and social effects.
#6 – The US government wants to expand their AI workforce.
After 36 pages of discussion, the R&D report concludes with two recommendations: 1) coordinate across federal agencies, and 2) create and sustain a "healthy" AI R&D workforce. Expanding the AI R&D workforce is a key theme across both reports as well as Chair Furman's paper. There are two ways to grow the workforce: 1) by training students at home, and 2) by importing talent from abroad. The reports are silent on how much emphasis the government should place on each approach. This depends partly on how long it takes to develop talent domestically. With an expanding set of resources for lowering the barrier to entry in the field (e.g., open-source software libraries for machine learning such as TensorFlow, Caffe, Theano, and MOOCS for learning machine learning online) some skills can be developed in short order. However, developing the PhD level expertise required to conduct research and make fundamental breakthroughs takes time. This level of talent likely needs to be aggressively recruited to the US in order to achieve the stated objectives with the urgency implied in the reports. Other countries that have significant pools of AI talent, most likely in their university computer science departments, and that have aspirations to participate at the frontier of AI-related fields, should brace for increased competition to keep their human capital at home.
#7 – The US government is not designing policy for general intelligence.
The reports define artificial general intelligence (AGI) as "a notional future AI system that exhibits apparently intelligent behavior at least as advanced as a person across the full range of cognitive tasks." The NSTC subcommittee takes the position that current policy should not be influenced by aspirations to achieve AGI: "The NSTC Committee on Technology's assessment is that long-term concerns about super-intelligent General AI should have little impact on current policy." The committee takes this position for three reasons: 1) many experts believe that AGI is not feasible in the short or medium run; 2) the authors assume the best way to prepare for AGI is to "attack risks" from narrow AI, such as security, privacy, and safety; and 3) the policy recommendations for AGI are unknown and may conflict with those for narrow AI, which is more certain and with immediate economic implications (implied). This approach to AGI, that it should have "little impact on current policy," is interesting because it stands in stark contrast to the views advanced by organizations that focus specifically on AGI such as the Future of Life Institute at MIT, the Future of Humanity Institute at the University of Oxford, and the Machine Intelligence Research Institute at the University of California at Berkeley. This creates an opportunity for other countries to differentiate by designing policies predicated on the assumption of achieving AGI sooner, creating an environment that is more attractive for AGI-related investment (e.g., regulatory frameworks that grant property rights and agency to machines).
Overall, these reports represent an important first step in the US government's AI strategy. The new administration will decide which of these recommendations, if any, it will follow. One of the recommendations calls for "The Executive Office of the President to publish a follow-on report by the end of this year, to further investigate the effects of AI and automation on the US job market, and outline recommended policy responses." If that recommendation is followed, then the resultant new report will set the tone, not only for the US, but for all OECD countries with respect to how they prepare to compete amidst this backdrop of technological advance and its effect on the workforce.
 This was assigned to the subcommittee on Networking and Information Technology Research and Development.
 The two reports are: 1) "Preparing for the Future of Artificial Intelligence," developed by the Executive Office of the President, NSTC Committee on Technology – Subcommittee on Artificial Intelligence and Machine Learning: https://www.whitehouse.gov/sites/default/files/whitehouse_files/microsites/ostp/NSTC/preparing_for_the_future_of_ai.pdf; and 2) "The National Artificial Intelligence Research and Development Strategic Plan," developed by the AI Task Force at the request of the Networking Information Technology Research and Development subcommittee of the NSTC: https://www.whitehouse.gov/sites/default/files/whitehouse_files/microsites/ostp/NSTC/national_ai_rd_strategic_plan.pdf.
 "Is This Time Different? The Opportunities and Challenges of Artificial Intelligence" by Jason Furman, Chair, Council of Economic Advisors, July 7, 2016: https://www.whitehouse.gov/sites/default/files/page/files/20160707_cea_ai_furman.pdf
 Interview on August 24, 2016 published in WIRED: https://www.wired.com/2016/10/president-obama-mit-joi-ito-interview/
Shared via my feedly newsfeed