1. Growing Pains as AI Enters a New Golden Age
Artificial intelligence is entering what many believe to be a new Golden Age. Machine Learning techniques, specifically Deep Artificial Neural Networks, has left the university research labs and entered the “real world.” Google acquired DeepMind for $500 million in 2014. Elon Musk, Sam Altman, and others pledged $1 billion to start the non-profit OpenAI in 2015. In 2017, Ford invested $1 billion in AI startup, Argo AI. Microsoft organized an AI lab with 5,000+ employees. Facebook started an AI Research lab. IBM is betting big that they can monetize their Watson platform. This is amazing to someone like myself who fell in love with AI when it was an obscure, far-fetched idea and machine learning was an emerging way to think about it.
The field of artificial intelligence (which includes machine learning as one of many tools in the toolkit) has changed dramatically in the last 3 years. Google, Facebook, OpenAI, and many others are in fierce competition over top talent, which can earn salaries in excess of $200k per year (or by some accounts more).
Companies investing in AI and machine learning have lured faculty away from universities (LeCun, Hinton, and Ng are among the most notable because of their work in deep neural networks), offering opportunities to field systems and products that can impact the daily lives of hundreds of millions of people. Moreover: Google DeepMind, Google Brain, Facebook AI Research Lab, and OpenAI offer pure research environments without some of the other responsibilities that come with faculty positions. In some ways, these industry labs can be more appealing than academic labs because of the availability of data and large-scale computing resources hard to come by (e.g., computing clusters). Where once being on the faculty at a well-ranked university was considered the pinnacle of success, industry now offers an appealing alternative. Newly minted Ph.D.s are turning to industry as a lucrative career path.
A perfect storm is brewing. Professors are leaving academia leaving gaps in university faculties to be filled. More and more graduating Ph.D. students are leaving academia, making it harder and harder to fill faculty vacancies. Meanwhile, enrollments in artificial intelligence and machine learning courses are skyrocketing. Universities are institutions steeped in tradition and are slow to respond to change. Some are waiting to see if these dramatic changes are part of bubble that will burst. Faculty hiring in AI and ML has been slow and cautious. In some cases — and perhaps counter-intuitively — it has been hard to hire AI, ML, and robotics faculty.
However, industry relies on universities to create the pool of AI/ML talent. Many in academia and in industry are worried that the AI/ML talent pipeline is fragile and cannot produce enough talent to fill industry needs.
My colleague, Jacob Eisenstein, expresses concisely what I have believed for a long time: if industry is worried about the AI/ML/robotics talent pipeline, they need to:
- Fund technology research
- Promote science funding and public universities
First, we will explore the issues that lead to concern about whether academia will be able to produce the talent needed for the industry using AI and machine learning. Second, we will look at the two paths toward easing concerns: industry funding of academic research, and industry support of science funding and public universities.
2. A Fragile Tech Talent Pipeline
Tech companies are concerned that the academic system cannot produce sufficient talent at a great enough pace to meet demand. As a consequence, there is fierce competition for those who can design and develop AI and ML systems. Google, Facebook, OpenAI, and others are competing for talent in the media and public relations realm, attempting to create a sense of primacy and inevitability about where the best AI and machine learning is happening.
The perceived successes in deploying and monetizing AI has created a demand for education in computer science, AI, machine learning, and robotics. Enrollments in computer science departments are growing rapidly. Enrollments in AI, machine learning, and robotics courses are increasing seemingly exponentially.
For university faculty to provide the best education, they must themselves stay current on a rapidly evolving field. One way to achieve this is for faculty themselves to be actively engaged in cutting-edge research. This is a good strategy; scientists can publish new research that feeds innovation and also keep their course material up to date. Furthermore, a secret of academic research is that professors learn from their senior Ph.D. students before sending them out into the world.
Demand for AI, ML, and robotics talent has outpaced the ability of universities to produce it. In part this is because university research is dependent on government funding and the growth of computer science has dramatically outpaced the commitment that the U.S. federal government has made to fund computer science research. The states have pulled back on support for public universities.
CS and AI academic researchers have had to turn more and more to industry for funding. This moves university research towards more applied problems. There is nothing wrong with that. However, applied research is not the forte of university research labs. Universities are really not well suited for work requiring industrial-grade engineering and the overall talent pipeline is degraded.
Meanwhile industry is pulling talent out of the system in the sense that fewer and fewer Ph.D. students are becoming professors. There are not enough educators for the growing demand and it is becoming harder to convince top Ph.D. students to stay in academia and become professors.
Further exacerbating the issue, talent is pooling at the most wealthy universities because they are in the best position to retain their faculty and to raise research funds to weather the storm. But even the top ranked universities can’t grow fast enough.
It is becoming more imperative that companies that have thrived in the AI, ML, and robotics spaces come to the assistance of universities to support the technology talent pipeline that they depend upon. As noted above it doesn’t take much to keep faculty productive in research and producing students. But it is better for industry to fund basic research for the reasons given above. One may even think of funding university research as a tax to support the talent pipeline (for the same reason we pay taxes to fund our public elementary and high schools).
I should note however, that to make the talent pipeline truly stable, the industry funding should be spread around to more than just the top universities. If industry funding only goes to the very top universities, then bottleneck will still continue to exist. Funding a broader spectrum of universities will help them build stronger degree programs and hire stronger faculty researchers and quality talent will come from more sources.
3. Why Invest in Academic AI/ML Research?
Nearly everything we currently know about artificial intelligence and machine learning has roots in university research laboratories. Companies that have benefited from academic AI research should re-invest a portion of their success in academic research. Even 1% of the $1 billion Ford earmarked for AI invested in AI research labs would have a profound effect on the academic landscape:
- It would ensure that faculty had the resources to mentor more undergrads, M.S., and Ph.D. students, helping to produce new talent. In the case of Ford, this talent would be likely be working on things Ford was interested in and can be hired immediately out of school.
- Universities would see the availability and accessibility of funding as a signal that it should invest in hiring more AI/ML/robotics faculty. This helps with the talent pipeline (but see below).
- Large influxes of funding into a research community can shift attention, focus, and research to new problems. Researchers that Ford is not directly funding may work on problems Ford is interested in solving for free (for hopes of acquiring funding).
- Some of the Ph.D. students will become professors. Most professors continue the work they started during their Ph.D. studies and their students will continue that work. Soon there are many more researchers working on a particular problem than before. Many of those researchers will get funding to work on those problems from other sources and publish papers giving their results away for free.
For companies like Ford, seeking to develop and deploy real-world technologies, a relatively small amount of investment can reap long-term benefit greater than the initial investment. This benefit is in terms of acquiring talent and surreptitious innovations.
Large influxes of funding into an academic community can change the direction of the entire community and have long-term inertia beyond the initial influx of funding.
3.1 Economics of Funding University Research
Why is university research so (relatively) inexpensive? The educational missions of those universities subsidize university research. Most tenure-track faculty receive 8–9 months of salary from the university and only need to raise 3–4 months of their own salary from external funding sources (NSF, DOD, NIH, industry, etc.). In exchange for this arrangement, faculty members teach and perform service to the university (i.e., help run the university).
Furthermore, university faculty may seek resources from government or other sources on related topics, allowing for value to be added to any research project.
Finally, some students — especially M.S. and undergraduate students — work for course-credit or for the experience of being involved in a research project that will make them more marketable once they leave the university.
3.2 Why Fund Basic Research?
When companies fund university research projects, they can choose to partner on projects that are narrowly applied to the problems the companies are most interested in, or they can provide grants without stipulation on what the funds are used for.
Universities are an economical means of producing (a) basic research, (b) research with societal benefit but no immediate profit potential, and © high risk research that is not guaranteed to succeed right away. Universities are suited for these types of projects.
University research is geared toward basic research over applied research for a number of reasons. First, the work will likely be done by Ph.D., M.S., and undergraduate students who are still in training. Second, major engineering and fault proofing is best done by engineers who are in the corporate chain of responsibility and supported by quality assurance staff. Third, university faculty and students will probably want to publish on the research.
University research is well suited toward research with societal benefit over commercial benefit. The incentive structure of faculty research is very different from the incentive structures in industry. Whether right or wrong, to receive promotion, faculty must publish on research results or show impact in the research community in other ways. Without the pressures of developing profitable technology, researchers can focus on problems that benefit society.
In artificial intelligence, there is a huge tension in industry between developing AI systems that can drive revenue and developing AI systems that are constrained by principles of ethics, morals, or privacy. Companies that pursue research in ethical AI must divert resources from profitable groups. Likewise, beneficial applications of AI may not yet have a market or may never have a market large enough to justify large-scale development. OpenAI is a curious counterpoint, which we will discuss below.
Universities are well suited for risky research. If a project fails to achieve the desired goals faculty do not lose their jobs. Therefore they can work on projects that might have a huge impact if successful but have a low probability of working perfectly. If the project fails, funding university research will probably have cost less than the equivalent number of full time staff that could have been working on something that was more likely to be successful. If the project succeeds, the company can continue to fund it or bring the research project in house.
It can be challenging to track how basic research translates into applied research. Research and development doesn’t happen in a vacuum, and applied work often takes inspiration from basic research. I admire companies that form basic research labs. Historically, research labs inside industry have had a poor track record of remaining independent from the corporate pressures to make a profit. Basic research labs often end up closed down or pressured to support current and future product lines.
4. The Role of Government
If stabilizing the tech talent pipeline is a tax on successful industries, why aren’t the corporate taxes already being paid by corporations supporting the talent pipeline? Government entities that fund university research, such as the National Science Foundation, are underfunded relative to the number of researchers seeking funding. States are underfunding their public universities. In other words, the taxes tech corporations are paying to states and the federal government is not being used to support the needs of the tech industry.
The U.S. federal government should be investing more in university tech research. It is good for the national economy to be a leader in technology. It is preferable that the next Google or Facebook is founded in the U.S. It is good for the U.S. for top scientists and top students to want to immigrate there. Finally, it is incumbent on the U.S. government to ensure that other countries do not gain a strategic military advantage over the U.S. because of technology.
Tech companies should pressure state governments to fully fund their commitments to their public universities. This will keep them healthy and able to handle the influx of computer science students. This will keep tuition down, which ensures that talented individuals do not need to be wealthy to get an education and enter the AI labor market.
Tech companies should pressure the federal government to increase the budgets of their funding agencies. Otherwise it will be incumbent for the tech industry to prop the tech talent pipeline up themselves at substantially greater cost to those companies. As noted above, universities are going to be more productive if they are not solely reliant on industry funding.
It is not unusual for large companies (or for an industry to pool together) to lobby state and federal governments. Typically, they seek favorable taxes or to influence regulation that might affect the industry. A lot of tech companies benefit from the university system. It only follows that the tech industry not overlook the symbiotic nature of industry and academia. Right now the tech industry is strong and would benefit in the long term for ensuring that high-quality education is available to as many people as possible in order to ensure that the high-quality talent they need will always be available.