You can see the full report here at jfgagne.ai/talent.
I reached out for help a little while ago on Twitter and on LinkedIn to assess the size and state of the global AI talent pool—a crucial issue for the entire industry going forward. Thank you to those of you from around the world who responded in large numbers. Your generous input has gone into a new report that we at Element AI have developed. We now have a more detailed picture of the size and characteristics of the pool of AI experts going into 2018. I see this report as a living document that will continue growing with others’ contributions. Our broadest measure of the global talent pool is 22,000 individuals: it remains clear that the fight for talent will continue into the foreseeable future. If you can help add more to this detailing of the global talent pool, you can reach me with the contact form or on Twitter.
Below are some of my observations on what I see happening around the world.
As AI is becoming a general-purpose technology, demand continues to grow for qualified people to tend the algorithms. Everyone in the world of digital technology development is looking at how they can learn the skills they need to work with AI. In every role—from researchers, to project managers, to software developers and more—demand far outstrips supply. The tiny size of the available pool of talent is actually holding back the AI wave, which would otherwise be far larger.
What we know about the talent pool is a blurry picture. The ballpark estimates offered by others range from 1,000 to 300,000, depending on the definition of talent. The 300,000 number, from a report by Tencent, includes the entire technical teams working on AI projects, whereas the 1,000 number counts elite researchers who have the 10+ years of experience necessary to lead major AI research efforts. Tencent says the demand as they define it exceeds one million, and it’s generally believed that all of the elite researchers have been hired into tech companies.
In our report, we look at another significant group, those who serve as the critical link between science and application. These are the engineers and researchers needed to develop the transformative AI applications made possible by the breakthroughs in the technology. A qualified AI expert needs to be more than just technically competent – he or she needs to have skills and experience that allow them to effectively mediate the various constraints of science, business, and industrial-grade software. When you think about it, it’s amazing there’s anyone out there at all with that intersection of skills.
At Element we had been working with an estimate that the pool for these experts was around 10,000 qualified people. That number was based on scraping the publication lists of the top AI conferences as a measure of who was influencing the field and who had the expertise needed to bring the new scientific discoveries from academia into application. We’ve since dug a little deeper to reflect the creativity in how people are filling these roles, and added some geography to our estimates.
Conferences Are Not the Only Proxy for Influence
Our own original estimate of 10,000 missed a number of influential people for two big reasons: we improperly limited ourselves to conferences as the only proxy for influence. Even that did not account for the language gaps that excluded many non-Western researchers. All we had to do was look around our offices and see that our search parameters for ‘expertise’ did not capture the reality.
The field of AI originally existed solely in academia. In the last ten years, however, we’ve seen most of the development shift from universities into the private sector, where the technology is solving real-world problems. The latest NIPS conference recorded only 12% of accepted papers as coming from industry, despite the migration. While some of the best AI talent is still found at academic conferences, our hiring process has identified incredible people in areas such as neuroscience or physics who are experienced in applying deep learning methodologies in commercial settings.
To count these other experts, we initially broadened our search to include people who have the needed skill set on paper (or in this case their LinkedIn profile) and a PhD in any field. This broad measure got us a total of about 22,000 profiles, with only about 3,000 of whom say they’re open for new opportunities. Demand is certainly outstripping supply: there are a projected 10,000 openings for AI development in the U.S. alone.
That said, it is becoming clear that the journey towards mastery of applying AI is similar to that for software developers of the early 2000s. It’s one thing to learn the language and maybe some of the latest tools and tricks, but a developer with ten years of experience has a ten to twenty times larger probability of success in creating the anticipated value of a project than someone who just recently finished school. After all, it takes a lot of time to become truly comfortable in that language.
Writing a book is an analogy anyone who writes can understand. It’s one thing to have a wide vocabulary and good grammar, but there’s a load of domain expertise one needs to acquire in order to start connecting interesting dots around which to conceive a book.
The field has a long way to go, both in continuing to make discoveries that will drive the technology forward and in supporting the needed talent development.
To sustain a good AI school or AI lab, you need skilled professors with at least five to ten years of direct experience. At the same time, the industry needs applied engineering and research—especially since the needs will become more niche and specialized as the technology spreads.
Though our measure requires a PhD, that’s not really true. Rather, the need is for deep technical expertise, and experience applying it is what really matters. This fact should make people reconsider talent pools with strong application efforts, even if they do not contain the most influential researchers.
In order to develop their talent pools, many up-and-coming countries have started investing in home-grown startups rather than attempting to draw in influential tech companies with deep pockets for research. In the report, if you look at the flows of talent in and out of the dominant US, you will see early signs of new AI hubs likely to challenge the incumbents in a few years. The largest migrations of talent are in and out of the US, but some countries have net-positive flows and many more are at par.
China is on another level: Tencent estimates that it is home to the second-largest number of AI experts and startups in the world. Much of China’s talent is relatively young, at least in terms of experience, with scholars who do not compare to North American professors due to China’s fairly recent pivot into AI in the past few years. But they aren’t shy, surpassing the US in publication volume and having a relatively high ratio of startups to researchers. Their “9-9-6” (9 am to 9 pm, 6 days per week) work/life balance will likely widen their lead in terms of volume, though it’s hard to say what the qualitative turnout will be.
Japan has some of the best expertise in the world, but it is heavily skewed towards an academic setting that is fairly weak at producing research in English. Japan’s researchers are also much more resistant to leaving their organizations, especially to join a foreign company, so it’s incredibly difficult to get access to this skilled pool. However, Anita Pan, a Trade Commissioner for Japan, cites a softening of Japan’s old ways, with large corporations opening up to partnerships with international players and decentralizing R&D through the startup ecosystem.
The United States and Canada have recently run into several problems nurturing AI talent. One is the brain drain: Big Tech has been poaching the talent from most universities, leaving few serious scholars to teach the next generation. Add to this the US’s 12% cut in science funding, and you can see that the US is pumping the brakes just as everyone else is kicking into high gear.
Rumor has it that China is making formal coding and the fundamental mathematics needed for working with AI mandatory from grade nine up. Widespread preparation for working with this new technology makes sense, given the importance of a digital skill set in today’s workforce. We will also need many bachelor programs to follow suit in order to support the myriad needs of the industry. As AI technology spreads, its long-term success will depend on increasing the diversity of mindsets able to work with it. (I was blown away by the response on LinkedIn with people sharing their ideas for introducing STEM education at an earlier age.)
While growth in Africa and India has been slower than in North America or Asia, these countries are in a great position to ‘leapfrog’ their current state.The very first AI Africa Conference in October of 2017 was a big success, drawing in expert researchers from all over Africa and the rest of the world to talk about real-world applications of deep learning. Their work also skews towards computationally efficient models and inexpensive sensors due to tight resources—developments that will be hugely valuable as AI scales.
The Future Talent Pool
As optimistic as we are about the possibilities of this new technology and about the available resources for democratizing it, it’s clear we must do a lot of talent pool development before we can realize its full potential.
We can see that a lot of the money in Asia and Africa is going to increase volume. But without deep research experience, combined with applied development, those investments are unlikely to be as efficient or impactful as they would be in North America.
No matter who you are, self-improvement is one of the most important and most overlooked attributes of young AI talent. It only takes four years of experience to become a senior AI researcher, or five years of experience to lead an entire institute. The determination and discipline to improve both the hard and soft skills continually will be the deciding factor in an AI researcher’s career.
This is true at the country level too. South Korea is the strongest Asian country at producing important English-language research, meaning their graduates are more eager to work with foreign companies than any other country I’ve seen. Opening channels to work more with global leaders in AI development could give them a leg-up in the global talent race.
In short, making the right investments now will make a huge difference in the long run, and we may very well see the geography of the talent pool shift around in a matter of a few years.
For the full report go to jfgagne.ai/talent.
Many thanks to all those from the community who reached out and contributed stats, anecdotes, translations and other useful information from their regions:
Bayo Adekanmbi | Ade Akin-Aina | Zainab Bawa | Adel Bibi | Valentine Goddard | Ian Goodfellow | Timnit Gebru | Ahmed Mamdouh A. Hassanien | Kiran Jonnalagadda | Jacques Ludik | Daniel McCormack | Michael James Milne | Shakir Mohamed | Anmol Mohan | Adeyemi Odeneye | Anita Pan | Rufina K. Park | Arjun Ram | Maged Shalaby | Yu Shao | Daniel Shinun | Ahmed Yousef