An interesting blog post by Jordan Schneider and Irene Zhang linked to an article in Macro Polo discussing the global distribution of AI talent. The graph below is based on the following definition in the Macro Polo article:
We created a unique and rich dataset of researchers with papers accepted at NeurIPS 2019, using that as a proxy for the top-tier (approximately top 20%) of AI research talent
Elsewhere they provide specific percentages for the undergraduate degree locations: China 29%, US 20%, Europe 18%, Canada 5%, UK 4%, Iran 3%, Israel 3%, rest of world 10%.
How should we think about those figures? Notice that the US has only about 4 times as much AI talent as Canada, despite having more than 8 times Canada’s population. On the other hand, the US has only about 2.5 times as many people of Chinese or Indian descent as does Canada. If these two groups are disproportionately represented in the AI field, it might explain why Canada is able to “punch above its weight” in this area. It seems likely that ethnic Asian researchers are also overrepresented in the undergraduate programs of countries beyond China and India, especially Canada, the US and the UK. If so, then although only 37% of top talent comes out of undergraduate programs in China and India, closer to 50% of global top AI talent may be ethnically Chinese or Indian.
Here is Schneider and Zhang:
NeurIPS was global. While most attendees sported American academic or industry affiliations, perhaps less than a quarter of the crowd had native English. My guess is at least 30% of the conference spoke fluent Chinese, with this set evenly split between Chinese and American affiliations. The rest were European or South Asian, with what felt like literal single-digit cohorts of Black and Hispanic researchers. The median age was around 28, reflecting how much new talent has come into the field in recent years. There was perhaps an 80/20 gender split.
To be sure, China and India have combined population of roughly 2.8 billion, but even so those ethnic groups may well be overrepresented in AI relative to their share of the world’s 8 billion people. It’s also worth noting that China is considerably richer than India, and also has a considerably larger share of top talent in the AI field. To summarize:
1. Asians appear overrepresented in AI, especially people of Chinese and Indian descent.
2. Income also seems important. Relative to population, Chinese do better than Indians, and Asians in North American universities seem to do better than Asians in Asian universities.
3. Japan is missing from the list, despite the fact that smaller countries such as Iran, Israel and South Korea show up on the list. Adjusted for population, the representation of Iranians and Israelis is even more impressive than ethnic Chinese and Indians.
The graph above also shows that talent tends to migrate toward the US after graduation. The Macro Polo article has another table that gives percentages of where these top researchers work today, which shows even further consolidation: US 59%, China 11%, Europe 10%, Canada 6%, UK 4%, others 10%. Based on those figures, Canada is the only country where the share of talent relative to population is even close to that of the US. (Perhaps Israel might also qualify.) This suggests one important benefit of Canada’s immigration policy, which has clearly attracted substantial tech talent.
[As an aside, I just saw the film Blackberry, which is an amusing look at a Canadian tech firm struggling to compete with American megacap firms.]
Nonetheless, it is hard for any country to compete with the US for both institutional reasons (our regulatory and tax laws are favorable to new tech companies) and the advantages associated with “network effects” (lots of AI talent collaborating in locations such as Silicon Valley.)
To summarize, tech talent first arises in places with the right cultural attitudes toward STEM education and the resources necessary to educate that talent. The talent then migrates toward the places where it can be used most productively—especially the US.
Ironically, this AI talent is engaged in a quest to produce non-human intellectual talent. Predicting the long run effect of that search is far above my pay grade, but I can’t help thinking the effects will be both unexpected and profound.
PS. Bloomberg has an interesting article discussing the wave of immigration that is occurring at the southern border. Apparently, it’s not just Central Americans intending to be farm workers:
The migrants are part of a growing number of the Chinese middle class on the run from an economic slowdown. They include entrepreneurs who saw business evaporate in the downturn, middle-aged fathers laid-off from China’s collapsing real-estate sector and young software engineers eager to make it in Silicon Valley. . . .
Data from the Department of Homeland Security shows the number of people with passports from mainland China crossing the US border without the proper paperwork has more than doubled in recent years. Almost 60,000 Chinese migrants have been detained for crossing the border illegally in the past 14 months, almost a quarter of them in California.
Lots of people believe the surge of immigration is due to the US having an “open door” policy. Lax policy might explain a part of the surge, but illegal immigration at the southern border also surged in 2019, a boom year. I suspect the actual reason is a strong labor market with relatively high wages (by international standards.)
READER COMMENTS
MarkW
Dec 21 2023 at 5:51pm
I suspect big regulatory advantages over Europe and China. Europe seems likely to impose rather severe restrictions on AI capabilities / uses (though not the UK, which punches far above its weight compared to the continental Europe). And China is likely impose strict limits for political / authoritarian reasons. It’s really hard to impose air-tight censorship on the sometimes ‘unruly’ output of LLMs — and if you clamp down hard enough, you may destroy a lot of the utility. Canada is kind of a wild card. It definitely has some of the European regulatory tendencies (‘a nation of hall monitors’) but open immigration and pretty easy collaboration with US companies.
AI talent is engaged in a quest to produce non-human intellectual talent.
I’m starting to feel like we’re in the ‘phony war’ period of AI. LLMs exploded on the scene. The things that these models can do are amazing, but the actual practical use cases still seem a bit elusive. Is the field going to find it much harder than expected to bridge that gap?
I myself have recently run into the Chat GPT ‘laziness’ problem. I asked it to do a task and it gave me a framework / example with the details left as an exercise for the reader, and despite repeated encouragement and cajoling, I couldn’t persuade it to finish the job. Since what I was doing was more of a test than anything, it was equal parts frustrating and funny, but if I was trying to get useful actual work out of it, it would have been a real problem. You’re the machine! You don’t experience boredom! Doing a complete job with some boring grunt work is why I’m trying to use you in the first place. C’mon, man!
Scott Sumner
Dec 21 2023 at 11:35pm
Yes, I think it’s an open question as to what LLM will be able to do. But it does seem like it’s only a matter of time before we have robots that can do much of the routine work that humans do today.
MarkW
Dec 22 2023 at 5:32am
Maybe. But I expect AI-controlled robots doing ‘routine’ physical work to be the very last thing LMMs do, not the first. There’s obviously much more money to be saved in replacing the efforts of highly-paid rather than poorly-paid workers. And LLMs have already show astonishing capabilities in a wide variety of intellectual and creative domains. Critically, no physical robots need to be designed, manufactured, powered and maintained to do these things. All of this means that LLMs are FAR more ready to do the work of writers, lawyers, artists, programmers, etc than they are to do the work of landscapers, hotel maids, roofers, or nurse’s aides. Even if LLMs could learn to carry out ‘routine’ physical work combining flexible movement in the physical world, manual dexterity, hand-eye coordination, frequent task switching, etc, it might not even turn out to be cost effective, given the high costs of the robot machinery. The intellectual work is the low-hanging fruit for LLMs, not the routine, physical work. It has turned out that the intellectual work that only intelligent, educated people can perform is computationally less complex than the routine physical labor that blue-collar folks perform (and the former, unlike the latter, requires no complex, costly, yet-to-be-invented machinery).
Scott Sumner
Dec 22 2023 at 12:18pm
Yes, I agree with that.
BC
Dec 22 2023 at 2:53am
If I understand the light gray flow paths correctly, it seems like the low-hanging fruit for the US to gain more AI talent would be to allow more USA-educated graduate students to stay in the US after graduation.
“Ironically, this AI talent is engaged in a quest to produce non-human intellectual talent.”
Or, maybe unironically and following well-established historical pattern, this engineering talent is engaged in the usual activity of producing more advanced machines, physical capital. The long-run effect might be the usual gradual growth in living standards.
Re: Chinese immigration at the southern US border. I thought the most common source of illegal immigration was visitors overstaying their tourist visas. The Chinese immigrants in the article needed to fly to Latin America before heading over land to the US. Is it hard now for Chinese to fly to the US on tourist visas?
Scott Sumner
Dec 22 2023 at 12:20pm
I believe it is hard to get a visa, but not impossible. Presumably these people were turned down.
Brent Buckner
Dec 22 2023 at 8:10am
I suspect that a significant determinant of the Canada to US ratio were idiosyncratic “bets” by Canadian universities on neural network AI research. Notably in the 1980s University of Toronto took on board Geoffrey Hinton, now somewhat hyperbolically referred to as “the Godfather of AI”.
Dylan
Dec 22 2023 at 8:59am
There’s also a high level of R&D subsidies, particularly in Quebec that have encouraged many international firms to locate their R&D centers there.
mira
Dec 22 2023 at 8:39am
I am likely (at least one of) your only regular reader who actually works in this field.
My guess is most of this is downstream of the US’ extremely strong capital markets, esp. for financing AI related things.
The lack of representation from Southeast Asia is maybe due to the fact that, unlike other fields, in my view ML/AI requires getting exposure to high-level calculus as soon as possible. I am not sure which Southeast Asian countries expose students to this in highschool, except I know that in China there are millions of students with vector calculus, etc exposure.
Scott Sumner
Dec 22 2023 at 12:21pm
I believe Vietnamese students are pretty strong in math. Less so for other SE Asian students.
Dale Doback
Dec 22 2023 at 10:02pm
Off topic, but personally I think AI is incredibly overrated. 15 years ago a lot of people were saying that drivers would soon be replaced by AI. We don’t seem close at all to replacing truck drivers with AI today. The LLM’s are impressive curiosities, but the suggestions that it could replace service professionals like doctors, lawyers, engineers, etc. anytime soon is even more of a joke than replacing truck drivers. I just don’t think we are in the midst of an AI revolution and I think this topic is prone to speculation and over represented in media. For example, I predict the various emerging GLP1s as weight loss drugs will have much more transformative effects on society than LLMs over the next 50 years despite being significantly less discussed.
MarkW
Dec 23 2023 at 6:09am
I was (and remain) an AI skeptic when it comes to autonomous vehicles. The real-time vision problem is very hard (especially in degraded conditions — rain, snow, fog — which is why all the test is done in places like Phoenix). And then, there’s no way an AI system can unload a truck or trouble-shoot and execute minor repairs on the road, nor predict/anticipate the behavior of other drivers, pedestrians, and animals alongside the road.
But in other domains, there’s no need for AI systems to interact directly with the complex, messy physical world. Nobody expects LLMs to replace MDs, lawyers, and engineers entirely, but they could well reduce demand as LLMs are used to replace/supplement labor in these fields in various scenarios. LLMs may prove equal to MDs as diagnosticians or, say, reading XRays and other medical images. Lawyers have already been caught out using ChatGPT to write briefs because of the the ‘hallucination’ problem, but that doesn’t make LLMs useless in this context, it just means you have to go to the trouble of double-checking and removing any made-up citations (much less effort than starting from scratch). LLMs are absolutely all ready to be used in music composition and performance as well and graphics arts. In those domains, even if they are used only to generate ideas and produce first drafts, they are immensely useful as is.
Dale Doback
Dec 23 2023 at 10:38am
LLM’s seem like they will be most effective when trained for specific tasks like interpreting X-ray images, but I am skeptical these sorts of highly specific applications will move the needle much.
Dylan
Dec 23 2023 at 12:00pm
I’m curious though, how much you’ve really played around with LLMs to come to this conclusion? There was a working paper published a couple of months ago looking at consultants at BCG that were randomly assigned to use GPT-4 or not, and the takeaway is that the ones that had access to it not only completed more tasks quicker with the AI tool, but the quality was judged to be 40% better than the control group without AI.
I’ve been using LLMs for a bit, and while I still have a lot to learn, I’m confident that even if the capabilities were frozen at the exact level we have today, there is room for huge gains in productivity just from people getting more exposure to the tools. In the last couple of months, I’ve used GPT-4 to make slides for a class I was teaching, analyze MRI images and give me a diagnoses before I got the official one from the doctor a couple of weeks later, take pictures of a record player I was trying to setup properly and give me specific advice on what adjustments to make based on the images and my description of the problem.
There are a lot more things possible, even with existing capabilities, once these are integrated into the tools we’re already using. Just the thought of having these capabilities integrated into Office…man …that would have saved years of my life when I was an associate. No more taking weeks to go through an Excel with a thousand companies on it, looking for the 50 that you wanted to shortlist for potential M&A. No more super late nights trying to make sure PowerPoint fonts are consistent throughout the deck and that all headers are pixel perfect aligned. Draft agreements to a much more professional level to start, minimizing the time that counsel needs to spend on them. Seriously, like 90% of my early career work felt like it was just formatting things, something that I’m not very good at, and which I’m happy to give over to an AI.
Dale Doback
Dec 23 2023 at 10:34pm
I’ve tried to use GPT4 for various things and found it underwhelming. It reminds me of how I felt about the ipad which offers almost no utility to me, as it is not portable like a phone but no keyboard or mouse either. Apparently lots of people use iPads and find GPT4 useful, just not me.
Dylan
Dec 24 2023 at 2:08am
Like most things, there is a learning curve with LLMs and it takes some practice to figure out what types of tasks they are good at and how to get the best results. Some of my students used GPT-4 (or 3.5) for their homework, and it was easy to spot because of how bad the responses were. But that is because they haven’t learned how to use it properly yet. It’s quite likely that other students used an LLM for an assignment without me even noticing.
I don’t claim to be anything close to a power user yet. I’m still working to change habits and integrate AIs into more of my workflows. And, a lot of what I’m doing now is just experimenting. I’ve likely “wasted” more time playing than I’ve saved from increases in productivity, but I expect that to change with experience.
P.S. I also don’t find the iPad all that compelling and have never owned one. Although my wife was gifted a hand me down one last year and finds it useful for making music on.
Thomas L Hutcheson
Dec 23 2023 at 8:01am
Imagine the benefit if the US went from discouraging migration of talent to actual recruitment! In some field it would also require regulatory changes.
Comments are closed.