China Will Overtake the US in AI, said Alphabet's Eric Schmidt
investopedia.com
investopedia.com
In another HN thread about AI, big corporations and China, someone mentioned that big American IT corporations are very active in China, offering their tech for pennies just to be able to test it on a wider scale, because in the West they can't do it (regulations, human rights, etc.).
There are many problems unsolved, where a lot of grunt work is needed and little progress has been made in the last decades - healthcare is expensive, many diseases are still uncurable, production is still cheaper to outsource to other countries than do it locally because automation (not just AI, also sensors, computing power, materials, etc.) are not there yet.
I think the general trend to see current AI iteration as the magical solutions to all problems is dangerous, as usual the reality is much more complex and should be tackled in a more thorough manner. Why not pay AI-specialist pay to get the best people into healthcare, new material research, etc?
The Chinese are making huge bets on AI, but they are beginning to sound like Japan in the 1980s with their Fifth generation computing project. Actually, there are so many analogues of China today to Japan in the 1980s (huge housing bubble, booming economy that doesn't seem to stop, a huge bet on AI....) that it doesn't really seem like a coincidence.
The difference is that ML today (at least the kind of technologies where China has and is gaining an edge) are deeply embedded in the commercial sector and public security.
We're not talking about China building a superintelligence in some kind of basement of a research facility, we're talking about facial recognition in public spaces, consumer products that get data from a billion customers who all pay and live on a centralised platform and so on. That's a reality already and it's not a stretch to say that China's environment is more suited for this than any other place on the planet.
I never said anything about neo tokyo, the fifth generation computing project was never vivid in our imaginations, all most of of us remember about it is the government throwing a bunch of money at AI research and hardware, and that it led directly to the 2nd AI winter. It also correlates nicely with other economic excesses (like China's problematic property bubble).
China still has a problem with attracting world-class talent, mainly because of their environment (great firewall, still way too wicked pollution). There are more Chinese working on ML in Silicon Valley then probably all foreigners working in tech in China. So if ML is a talent game, then I don't see how China will be able to win that in the near term, even with all their local human resources.
And on the data and policy side China has hugely favourable conditions. I think we're making a great mistake if we're over-hyping the role of talent.
For instance, there is no amount of cluster power that'll let you best resnet with a non-residual network. Well perhaps there is, but even Google's mighty clusters can't make up the difference.
In audio, end-to-end models that take 2 weeks to train on a dual GTX 1080 machine, and take > 1 week to train on an entire datacenter beat everything we've ever come up with in linguistics. And it was recently shown that taking the convolutions and redoing in hand-coded assembly gains you more performance than the quite-meager performance a cluster gets you.
The same goes for data. We can outperform VGGNet trained on all of imagenet (~ 15 million images) using resnet trained on a 1% random sampling of imagenet. We have papers putting forth statistics on specific model architectures outperforming with < 0.1% of data of benchmarks and just-barely-worse performance using a decent, but fixed and tiny number of data.
I understand why Google and Microsoft are pushing "you can't beat clusters", but anyone fiddling with batch-size on actual models will quickly realize that it's ... let's say "riddled with MAJOR caveats" (using clusters implicitly forces large batch sizes, which costs you a LOT in training efficiency, especially once you get past the first few rounds, and the better your model, the more it costs).
okay granted, but isn't this largely part of the academic field, openly accessible and transferable? People at Baidu and Tencent are likely up to date on the state of the art.
Contrast this for example with the American military edge or the economics of the German Mittelstand. There's a degree of ingrained knowledge that gives countries decades of advance, but academic research can often be transferred. Understanding in ML to me seems to fall mostly in the second category, and when that category dominates then data and policy really does become relevant.
I'm not sure we have seen an ML intellectual arms race yet. For the most part techniques seem to spread fast and as a result data functions like oil of sorts and gives the biggest players the largest advantage.
1) they actually have an older implementation. Not many people have working linguistic voice recognition, as that's really, really hard, and you're just not going to make one of those without a 100-strong multidisciplinary team. There's no real business need to replace it (yet).
2) once you have a multidisciplinary team, the multidisciplinary aspect of it is the main reason for the size of the org. Moving to a pure-AI solution would mean 80% or so of the department would become useless and unable to contribute.
3) I bet the linguistic model looks a whole lot more comfortable to the executives than a pure AI algorithm. After all, a linguistic model will not make "hidden mistakes" (mistakes that the system makes but have never been programmed in). And let's not forget that the last few hidden mistake in a highly public model was confusing African Americans with, shall we say, animal byproducts. Needless to say, this was NOT good PR-wise.
(by the way you should try to have a realistic "let's do this with AI" talk with a senior manager, and you'll see what I mean. "Can you guarantee it won't make mistakes ?", "Nope. In fact I will pretty much guarantee mistakes. It's like a person. It might purposefully make mistakes in the sense that it causes a disaster in one area because it improves it's metrics". "Okay. Can you at least tell me why it made decisions ?", "No. Impossible. Also: please don't believe any AI researcher claiming otherwise". You're asking extremely risk-averse people to take a big leap)
4) Career-opposition. In the large orgs, the senior engineers have their senior position because they improved step 57 of algorithm 21 by 5%. Making proposals to replace everything after step 3 with an end-to-end model ... they will "politely sabotage" it. (e.g. demand guarantees that it won't make mistakes. Request papers proving that it outperforms, not just the function, but every individual step. Demand they illustrate that thousands of slight mistakes won't ever happen, ...)
(you know, like factory workers demanding robot features BECAUSE they figured out that they're ridiculously hard. They have no use for the factory. E.g. demands that a robot responds intelligently to a human walking by ... on a floor where no humans are allowed if any machinery is running. Or demand physical separation of robot action radii, when the software supports those robots working together and this is in fact used in said production line. Makes no sense whatsoever. What are they doing ? They think they're defending their jobs)
I've seen several startup products based on Deepmind's Wavenet, but Microsoft and Google's voice recognition have both said that theirs is based on a linguistic accoustic model. You know, the huge, very very complex, dozens of different components, each their own specialization.
So the opposite of what you'd think is actually what's happening. The organizations doing the cutting edge research are not ahead of the curve, they're behind, far behind, and falling more day by day.
Or to put it another way, you want to see innovation happen ? Find companies that would be dead and bankrupt without that innovation. Google and Microsoft, those are not it. Even facebook is better.
I don't really believe that the implementations out of the startups are meaningfully/significantly better than the implementations out of Google/Facebook. If they were seriously better they'd be acquired (as we've seen over and over again).
* 9 out of the top 10 algorithms for the Stanford Question Answering Dataset are by Chinese teams, backed by HIT & iFlLYTEK, Alibaba, Microsoft Research Asia (in Beijing), NUDT & Fudan University, and others. They beat out teams by Samsung, Salesforce, CMU, and Facebook, for example.
https://rajpurkar.github.io/SQuAD-explorer/
It is not like others do not try. This is a highly commercializable area of research and there is no requirement to publish the technical details of the submitted solution.
* Based on publications from 2015 onward, Peking and Tsinghua University are ranked in the world's top 5 for published papers in top AI conferences (at #3 and #5).
http://csrankings.org/#/fromyear/2015/toyear/2018/index?ai&v...
See my previous comment for more info & counterpoints about talent and pollution.
There's a lot of people publishing ML work on public datasets which barely move the needle. It's usually not too hard to squeeze something extra into a model to do better than the previous work, and if not there are always other datasets, so these raw numbers don't really means lot IMO.
Not to say China won't overtake the rest of the world, eg lots of great work coming out of MSRA, but I think that there is a lot of truth to the idea that once you start measuring something, people start gaming it, and I think the political environment in China encourages this gaming, so I would pay less attention to metrics like this that can be gamed.
Even paper metrics are gameable. Most reviewers acknowledge that once you remove the obvious accepts and obvious rejects you've got a large pile of passable but forgettable work. So if you can just keep cranking out passable work that fits the zeitgeist, you'll get more papers accepted.
I might be wildly off base here, I haven't tried to construct statistics, and academia has a large amount of starfucking going on, so it may be a cultural bias thing, but I don't get the impression that China is anywhere near dominant yet, but yet all these statistics look great!
Part of their 2030 plan to become preeminent in AI is to develop indigenous breakthroughs by 2025 and be equal to the leading nation by 2020. They are very much on track to achieve the closest goal given evidence to date.
With that out of the way, all of those are not very relevant. Its like judging the quality of programmers by ACM programming contest wins, or by judging hardware advancement on super computing benchmarks that no one really cares about anymore.
China has declared that they will be the winner in AI in the future. But they definitely aren't today, and that assumes that everyone else stays absolutely where they are already.
As shown by PISA and other international tests, the percentage of AI-capable people varies wildly among the populations in that 7.4 billion global pool. China performs very well, esp relative to their development level, and above most OECD countries.
http://ncee.org/2016/12/pisa-2015-results/
If China can hold onto most of its talent, it would be a relatively large percentage (say 4-5%) of 1.4 billion vs a smaller percentage (2.0-2.5%) of 1.3 billion (OECD country populations) and a very small percentage (say 0.2-0.5%) from 4.9 billion (the rest of the world) who can potentially work productively in AI. This assumes the Gaussian distribution of education quality, skills, and perserverance and a fairly high bar (2+ SD above OECD average) for a lower threshold to work competently in AI.
It is obvious that not all talent in the latter two would like to immigrate to the US, so it is unclear if the US will continue to have a quantitative advantage despite a more global talent pool. Chinese government's deliberate policy in nurturing and developing AI talent will also add to the number of local practitioners. There is also clear evidence that more Chinese nationals are moving back home after graduating from abroad.
The rough estimates above are relevant for AI optimizations and applications in many real-world settings since a sufficient quantity of R&D practitioners is necessary for that.
For breakthroughs, other factors might come into play and it is not clear if cultural factors and upbringing would affect the potentials of people to create paradigm-shifting ideas, which means quantitative analysis in isolation is not nearly as predictive.
The number of Chinese returning after studying abroad is increasing because the numbers studying abroad are increasing, these are mostly rich fuerdai kids not studying STEM who have no chance getting jobs in the west anyways; the amount of STEM talent retained, especially at the important post graduate level, has remained fairly constant.
Overall, China cannot achieve dominance in this area with its current policy of nationalistic tech isolationism (e.g. blocking off internet sites, not allowing immigration). Even if we ignore the west, there is still India and other countries that have large roles to play; that have their own promising characteristics.
Granted, they are likely above China's average in terms of skills but I read elsewhere that someone involved in PISA said regarding about unofficial PISA 2009 research in other Chinese provinces that even some very poor regions scored close to OECD average.
We will likely see China's whole country results in 2019 when PISA 2018 official results come out and I believe it will still be better in math than many OECD countries.
Your point regarding lower average quality of education in China could be true but that means in 10 years they could be improved with economic development and their skills would be even better than now. Developed countries in East Asia and Singapore where populations share the same cultural traits with China have for some time dominate PISA rankings in all subjects.
India has many talented people but because of many factors, as a percentage, the figure is probably many times lower than China. Check out their former PISA results before they pulled out. They will improve but it is a long way to go and we know from history that network effects and other factors tend to make tech dominance into natural monopoly, at least until a major paradigm shift happens.
No other country has both the size of native population and the distribution of population's skill levels taken together to compete on equal footing with China in the next 1-2 decades.
The US can hold its own and perhaps keep the lead by being a desirable place for global top talent but it remains to be seen if the government's policy and other factors will erode that advantage away.
Then we’ll need to lower the number of total population used to calculate AI-capable natives somewhat. Given data on primary vs secondary school enrollment, the adjustment ratio is about 0.65. That would adjust the relevant population size down to about 0.9 billion.
This has been an interesting exploration and shows how hard it is to predict the future given so many factors involved. I still believe the broad-stroke prediction that China will lead in AI-based technologies in 2030 will still come true though.
What happened was that advances in scalar processors completely blew the specialized processors out of the water, to the point that general-purpose hardware could implement the desired programming features faster than the dedicated hardware. On top of that, the expert system programming model turned out to be a massive bust. In effect, Japan was suddenly faced with expertise building machines that solved problems no one wanted to solve, and were inferior even at that.
I think that the same risk is equally valid today: it's possible to so focus on one particular aspect of AI that, if the programming model ends up being revolutionized, you're left with useless hardware.
Because that's expensive, and companies are still structured in a manner built and optimized for repetitive manufacturing. They are not built to, and as a result can not, value things like knowledge and skill. Everything from why they have offices, to how those offices are built, to how the organizations are structured is anathema to having a person or persons with valuable domain knowledge as a key component. Read essentially any book on business management written since 1980 up to today. Every single one of them will advise managers to highly prioritize the dispensability of employees and inherently emphasizes accounting ('business') oriented roles over technical ones.
Remember, IT is a cost center. Anything spent on it, or those with skills in it, is pure loss that offers no value. You're a healthcare company, not an IT company, and IT people don't even see patients.
In the last 5 years, participants from China reached the Google Code Jam finals (top 25 contestants) 12 times. Russian participants did the same 40 times (notably last year there were 13 Russian contestants, half of the finalists).
If this is evidence that China is going to dominate AI, what does it say about Russia?
The argument about AI lead is based on other facts, not particularly this one.
It seems to me that people who lack other means of signalling excellence and are ambitious to become visible internationally are most likely to enter such competitions.
What will kill the US in AI research is when the government starts slapping export controls on everything and preventing foreign nationals from doing top level AI research in the US. Then all the talent will go to other countries.
https://www.ft.com/content/856753d6-8d31-11e7-a352-e46f43c58...
Dear Alphabet,
Please make your products and services work before worrying about geopolitical balance of power.
What issues do you have with spam? Personally, I get maybe 1 false-positive spam email per year with Gmail, and maybe a few false-negatives (which are less of an issue anyway), for an accuracy rate in the 3-4 9s range. So I'm intrigued as to what you're doing or where you encounter the inability for gmail to distinguish between real communication and spam.
The other issue I have is that the search function doesn't update... I try to search for an email from the day before and it won't show up.
(presumably because google doesn't like 'free' servers)
Edit: my ideal UX in this situation would be to get the mail in inbox, with a small notice saying "Unverified" and a mouseover/hover text explaining what that means re: SPF records; from there if you mark it as spam it would treat such unverified mail from that domain as spam on an ongoing basis
I've also noticed recently that spam hits my inbox before being filtered, which is aggravating as that means I get a notification about it.
- I can't find any concrete data, but it feels like at least 50% of publications on conferences like NIPS or CVPR are made by Chinese nationals. Even if it is initially from USA universities, many researchers are going back, either due to unfavorable Visa situation or else.
- Even among Silicon Valley investors, funds with China roots are relatively much more into AI startups
For all its faults, their leadership has their eyes properly set on science and technology as a means to challenge the US economically and militarily. They're incentivizing the proper areas where they want industry to operate, and are better poised to take advantage of it. The US government is hampered by dysfunction in comparison.
Would hope the US sees what's at risk here, and moves from underestimating the threat, to overestimating it and taking massive action as they often do.
Better summary here: https://worldview.stratfor.com/article/coming-tech-war-china
https://www.theatlantic.com/technology/archive/2017/02/china...
> The 2017 AAAI meeting—which ultimately relocated to San Francisco—wrapped up just last week. And as expected, Chinese researchers had a strong showing in the historically U.S.-dominated conference. A nearly equal number of accepted papers came from researchers based in China and the U.S. “This is pretty surprising and impressive given how different it was even three, four years back,” says Rao.
* 9 out of the top 10 algorithms for the Stanford Question Answering Dataset are by Chinese teams, backed by HIT & iFlLYTEK, Alibaba, Microsoft Asia (in Beijing), NUDT & Fudan University, and others. They beat out teams by Samsung, Salesforce, CMU, and Facebook, for example.
https://rajpurkar.github.io/SQuAD-explorer/
It is not like others do not try. This is a highly commercializable area of research and there is no requirement to publish the technical details of the submitted solution.
* Based on publications from 2015 onward, Peking and Tsinghua University are ranked in the world's top 5 for published papers in top AI conferences (at #3 and #5).
http://csrankings.org/#/fromyear/2017/toyear/2018/index?ai&v...
* China plans to add AI and programming courses in primary education. It is likely that soon, if not already, the proportion of US middle and high schools that offer computer science courses will be lower than China's.
http://en.people.cn/n3/2017/0828/c90000-9261282.html
* Most Chinese kids spend a significant amount of time working on their academics, even outside of school (whether this is good or bad at the individual level is different from the main topic of discussion here), while the same cannot be said about most American children.
* Since 2000, China's team was no. 1 in International Math Olympiad 12 times (out of 18) and 5 times at no. 2
https://imo-official.org/country_team_r.aspx?code=CHN
* Mathematics is an important foundation of AI. One needs to be relatively competent at math to understand most AI research papers and even textbooks. The level of math required may look basic to some HN readers but it still takes many years of focused learning to master for most people. Many smart people do not master math simply because they haven't spent the time.
The difference in math skills between the US and China high school students is evident in PISA 2015 results. My back-of-the-envelope calculation suggests that the potential proportion of AI-capable natives in China could be 5 times as large as the US. Since China has more than 4 times the population, the potential number could be 20 times.
https://news.ycombinator.com/item?id=15569339
* China is doing everything it can to attract back their top talent as well as those from other countries. A reason for pollution curb, which starts to become successful, is to satisfy talent's requirement for quality of life. They have instituted a competitive program to attract global talent. (At the same time, the US is not exactly sending a welcome signal.)
https://en.m.wikipedia.org/wiki/Thousand_Talents_Program_(Ch...
https://www.nytimes.com/2017/05/27/technology/china-us-ai-ar...
It's instead the encouraging atmosphere and resultant inflows of capital (intellectual and otherwise) that may put China ahead in the future.
AI requires more than math, but one cannot work competently in AI without being competent at math. Based on teaching experience, I would estimate the lower threshold to be about 2SD above global average (or equivalent to 700 PISA score in math), which might mean less than 2% of US high school students are above that threshold.
Sure, I can certainly get behind greater funding for scientific research and education for our benefit, but China is certainly not a country I would like to emulate whatsoever.
My biggest question is, what is the meaning and implication of China "dominating" the AI industries? China overtaking the US in military and commercial might? This honestly seems unlikely.
And a follow up question is, besides government funding, what else is China doing to secure AI dominance? Putting human rights and regulations on the back burner would appear give them a competitive advantage - but is this worth the cost of privacy and freedom for the average citizen?
'"Shockingly some of the best people are in countries we won't let into America,' he said. "Iran produces some of the smartest and top computer scientists in the world. I want them here. I want them working for Alphabet and Google. It's crazy not to let these people in."'
Well now I am not following this logic at all. What is China doing to make working there more attractive? I would imagine the US looks a lot more attractive than China in terms of freedom and safety regulations.
If anyone has a copy of the AI report in English I would greatly appreciate it.
Calm down, it hasn't even reached 1.5 billion yet.
That means China will dominate in most engineering fields. Also some where the US today is an absolute leader.
Economic figures from China are not very reliable. But they are making smart investments while the US is mostly moving money from one pile to another.