Plus, What Peter Thiel's NYT oped gets wrong about US-China AI competition (hint: a lot)
|Aug 5 at 3:58 am||Public post|| 1|
Welcome to the ChinAI Newsletter!
These are Jeff Ding's (sometimes) weekly translations of Chinese-language musings on AI and related topics. Jeff is a Rhodes Scholar at Oxford, PhD candidate in International Relations, Researcher at GovAI/Future of Humanity Institute, and Research Fellow at the Center for Security and Emerging Technology. Check out the archive of all past issues here and subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (those who can pay support access to all content for all).
*The Google Doc link (click heading) is almost the same as the newsletter body text for this issue (though there might be a few additional annotations on the Google Doc) - I’m including it mostly as a space for people to add comments and ask questions
Longtime readers of ChinAI will know that one of my favorite sources is Leiphone. This Chinese new media platform, founded in 2011 and based in Shenzhen, has emerged as a leading portal for science & tech coverage. Past ChinAI issues have covered Leiphone’s fascinating interview with the head of Baidu Security on the “Long Front” of AI cybersecurity, its AI Impact Factors Database which aims to track every paper publication/ competition result/ development project/ corporate activity in the AI field, and Huawei’s move into the AI+Security Industry.
Instead of featuring a translation of one article from a random Chinese site that most readers will forget by the following week, I want to change things up and have us browse Leiphone’s home page together. We’ll translate some excerpts of interesting articles along the way but the primary point of this exercise is to use Leiphone to give folks a sense of the scale/depth/breadth of Chinese S&T media coverage — it’s a nice reminder that the translations featured in ChinAI are just a drop in the bucket. Okay — let’s browse together!
Screenshot of Leiphone’s home page (https://www.leiphone.com) taken at 9:30AM August 4, 2019 (US Central Time). I’ve divided their home page into five sections, which we’ll go through in order. Formatting note: everything that’s a direct translation of Leiphone content is italicized; everything else are my annotations.
1）Leiphone: Understanding Intelligence & the Future (Header/Navigation Bar)
First row tabs: AI Research Institute; AI Investment Research Group; Activities Center; Top Academic Conferences (new!); Special Topics; Love Playing with Machines
Second row tabs: Industry, AI, Smart Driving, AI+, Finance Science and Tech, Future Medicine, Cybersecurity, Smart Cities, Robotics, Industrial Cloud, Smart Hardware, Internet of Things, Global AI and Robotics Conference)
The links in the first row of the header/navigation bar all go to separate branch sites. For instance, the "Top Academic Conferences" tab goes to an entire portal dedicated to covering academic conferences and articles/coding tutorials/open source software updates; The "Love Playing with Machines" tab goes to a different branch site for folks interested in playing with smart hardware products.
The links in the second row keep you within the main site but just filter the articles by the topic. Note the last tab (GAIR) — a relatively well-known AI conference in China — which Leiphone has hosted annually since 2016.
2）Industry Information (Vertical)
“Didi Chuxing open sources ‘Delta,’ a training platform for natural language understanding model | ACL 2019
“Douyin’s On-line Group Chat Function Combines Forces with Duoshan to Check WeChat”
“IDC Q1 Report on China’s Cloud Services: China becomes the World’s Second Largest Public Cloud Market.”
“Rootcloud CTO Zhen Liu Leaves Position, the Rivers and Lakes of the Industrial Internet are Turbulent”
This section has mostly short (5-6 paragraph) pieces on industry trends. Some, like the article on Didi’s new training platform, read like edited press releases. But others, such as the last one on Rootcloud, contain some good analysis. Here’s the last paragraph from the Rootcloud article: “According to Leiphone’s understanding, although startups in the industrial Internet field are constantly emerging with hundreds of millions of financing, there is still truly demand for digitization from manufacturing enterprises (that is not being filled). However, the entire industry is still in its infancy, and the business models of many industrial Internet companies are not clear enough. Their technical capabilities are unable to meet the fragmented needs of the digital transformation of manufacturing, and there is also a big bubble in the capital market. There is still a long way to go before the manufacturing industry really enters the industrial Internet era.”
Under the pressure of the United States, Huawei shipped 181 million mobile phones in the first half of the year. Hongmeng will soon debut
2019 10 Top-level Python Libraries you have to Know
Research on micro-expressions using deep learning: difficulties, progress, and trends | CNCC 2019
Complete Hardware Guide for Deep Learning
2019 China National Computer Congress
These five slides functions as links to featured articles. Interestingly, the complete hardware guide one goes to one of the separate branch sites (Leiphone’s AI Research Institute portal); the last one goes to the site for the CNCC.
4) Featured Longreads
“Explaining the Global Smartphone Market: Can Huawei Replace Samsung as the World’s Number One?”
“Apple’s Payments of $4.7 Billion did not Stop Qualcomm’s Plunging Stock Price”
“Putting NB-IoT Under the 5G (umbrella) Expands the 5G Vertical into a Very Vast Road.”
“30 Years of Technological Accumulation, More than 300 Patents, this Chinese female AI scientist Chooses to Return Home and Start from Zero.”
“Sony & LG’s Mobile Phone Business: The Former Kings, Now in the Bronze Age.”
“DingTalk’s Hospital of the Future at the End of One Year: How to Build a Hospital’s Digital Base?”
This section is pretty incredible — all written within the past two days — all 4000+ Chinese characters in length (6000/7000 word range if translated into English). Again, each one of these would be candidates for a weekly ChinAI feature translation and each would deepen/widen our knowledge pool about China’s tech scene — the knowledge arbitrage in this space is incredible and I would encourage people to take the ChinAI model and run with it
Let’s take a look at an excerpt from the article (link to original Mandarin) about the Chinese female AI scientist:
Context: Pensees-AI, a company that focuses on computer vision, IoT tech, and industrial applications, announced the opening of a research institute in Singapore, with a technical committee that includes a broad range of Singapore professors (including the former VP of the National University of Singapore). Pensees-AI also signed a memorandum of cooperation with AI Singapore (AISG), a national AI programme launched by Singapore’s National Research Foundation (NRF). The bulk of the article focuses on Shengmei Shen who will lead the Pensees R&D institute in Singapore. She who was formerly the Assistant Director of Panasonic’s R&D Center in Singapore, where she had worked since 1992. Her team won more than a dozen top-level international competitions in computer vision, including the US NIST IJB-A facial recognition challenge and Microsoft’s MS-Celeb-1M Challenge.
Direct translation of some passages about Pensees-AI’s current Singapore research team — which reflects how Chinese companies (unlike Peter Thiel) seem to get that diverse, global teams produce results:
At present, there are nearly 30 people at Shengmei Shen’s Singapore Institute, which is located in the town of Shenzhou, and they come from more than a dozen different countries. Shengmei Shen dubbed the team “The United Nations."
Although the team's "United Nations" attributes come from the missions of multinational companies intermingled with Singapore's geopolitical position, she believes that this background is very beneficial to team building and scientific research. First of all, people of different backgrounds have different advantages in ideas/thought processes, and team members can form beneficial complementarities. In addition, the cohesiveness of the "United Nations" team is not based on the same cultural background but rather on the same values, and it is easier to generate deep mutual recognition.
As for role this "United Nations" will play, Shengmei Shen revealed that at present, it is necessary to base (their work) on the product and business model of Pensees, and to integrate Pensee’s development direction with R&D and innovation.
5) Recommended Topics and Industry Express (News)
First, the recommended topics section includes promotions for major conferences and events. When I took the screenshot, the slide featured the World Artificial Intelligence Conference, which takes place in Shanghai later this year.
Second, the Industry Express (News) section is similar to Industry Information Verticals section. The two articles in the screenshot:
“Alibaba Cloud takes top three in the world’s container market, first in China, and enter into the quadrant of strong performers.”
‘Ant Financial’s Xiandong Jing: If Xiang Hu Bao works well then the insurance industry will be better.”
Maybe the main difference between this Industry Express (News) section and the Industry Information vertical is that this section summarizes work from other publications whereas the Industry Information vertical features mostly original Leiphone work. For instance, the interview with Xiandong Jing, CEO of Ant Financial, came from the just-published cover article in the magazine Chinese Entrepreneur.
Debate Segment: Thiel’s Messy Spiel
Somehow The New York Times published a Google hitpiece/Palantir marketing pitch by Peter Thiel as an oped last Thursday. In a piece that meanders from questioning Silicon Valley’s “cosmopolitanism” (which for some reason is put in quotes) to critiquing those who are worried about AI’s risks for all of humanity to defending a zombie “Cold War mentality” against China, Thiel’s main argument is that Google is helping the Chinese military instead of America by opening an AI lab in Beijing.
Thiel gets SIX major things wrong (and counting! See my Twitter thread for some good back and forth on some of these) in his oped:
1. He gets basic facts wrong: Thiel argues that China's constitution “mandates that all research done in China be shared with the People’s Liberation Army." This is just not true. As Lorand Laskai outlined, military-civil fusion certainly does incentivize research-sharing and “aims to create a commercial market for private firms to compete for PLA contracts” but Thiel’s claim takes this to an unfounded extreme.
2. No disclosure of conflicts of interest: In the piece Thiel hypes AI's potential to help armies gain an intelligence advantage (that might have been an appropriate time to mention his company Palantir won a decade-long, $876 million contract to do just this for the U.S. army last year or that Palantir has at least 29 active contracts, worth a combined $1.5 billion, with the U.S. federal government). As @ConMijente pointed out, this read like a marketing pitch targeted at government officials who dole out billions in defense contracting.
3. Thiel has a very confused conception of AI: He first says that AI is a military technology at its core. He claims that, as was the case with nuclear fission, “the first users of the machine learning tools being created today will be generals rather than board game strategists.” Um what? Never mind the fact that there are already many first adopters of machine learning tools across a wide variety of commercial verticals (translation services, predictive services in finance, etc.) You could actually take Thiel’s statement and completely reverse it and it would be true. The first users of the machine learning tools being created today will be board game strategists rather than generals. Top chess and Go players use machine learning-backed engines to improve; top generals are taking steps to make sure machine learning tools are robust before adopting them into mission-critical operations.
4. Thiel can’t be this ignorant about AI; it seems like he is deliberately trying to give an ambiguous conception of AI to take a shot at those who are concerned about the risks posed by AI for all of humanity. He waffles back to saying AI is dual-use in the middle of the piece and then says this ambiguity is "strangely missing from the narrative that pits a monolithic 'AI' against all of humanity." It’s hard not to take that as a direct shot at my home base, the Future of Humanity Institute, directed by Nick Bostrom who wrote Superintelligence (which warned about the risks of artificial general intelligence). Thiel is randomly attacking a straw man in the “terminator” AI meme. As one of OpenAI’s earliest backers, Thiel should know that there are more nuanced views on AGI’s arrival. Take for example: this reframing of superintelligence by FHI's Eric Drexler, who is widely regarded as the founding father of the nanotech field, in which he outlines a trajectory toward comprehensive, superintelligent-level AI services.
The broader point here is that Thiel seems to not get the basic idea that AI can be many things at the same time. We can recognize that AI like other general-purpose tech (e.g. steam engines or electricity) can empower both civil & military applications AND also see that intelligent agents pose unique risks. The risks of AI exceeding human-level intelligence are just one subset of unique risks posed by AI (others include accident risks from increased automation), but it's definitely not to be dismissed. A wide range of AI experts take it seriously.
5) I'd argue Google's efforts to open AI labs in China aren't "cosmopolitan"/anti-US but lean more toward being self-serving/good for US innovation given global talent flows. Thiel and many others who think that US should not be involved in any offshore R&D in China ignore insights from vast body of lit on the globalization of innovation and tech flows. Let's take a look at one of the key papers from this lit (Eaton & Kortum 1999). Drawing from international patent data from the five leading research nations at the time (US, Japan, Germany, UK, and France), Eaton and Kortum find that 40% of U.S. productivity growth came from research performed in the four other industrial leaders. They also show, through a counterfactual experiment why tech isolationism would be such a stupid tactic for the US to adopt: "cutting off the United States from the rest of the world would cause its productivity to fall far behind the other four."
The lesson here is that Google, other tech companies (e.g. Microsoft Research Asia in Beijing), and 1000s of MNCs w/ R&D labs in China aren't doing this work out of charity or some deeply-buried desire to help the Chinese military; rather they want to be plugged into global innovation networks and adopt tech advances from abroad into home bases. Now, are Thiel and others right to point out some of the negative externalities (e.g. indirect leakage to enable some Chinese mil. developments, building up talent that move to Chinese competitors who could overtake in the long-run?) Sure, but let's have an open debate with real arguments backed by empirics instead of ad hom attacks on the patriotism of companies like Google.
6) In one respect, I agree with Thiel. There’s one strand of the zombie “Cold War” mentality never seems to die: the old playbook of leveraging the exaggerated fears of Cold War competition with a rival in order to advance a totalizing technocracy that wields complete control over society. The historian Walter McDougall’s warning, issued about the dangers of America’s post-Sputnik techno-nationalist turn, still rings true today:
The social mechanisms required to tap the full technological potential of a nation, particularly in the context of cold war competition, mean we have to pay a price for our advances in science and technology and the price is usually a sacrifice in human values. I believe it is inevitable, as long as international competition is the primary engine moving history, and technology is brought to bear in the competition, that we will move more and more toward management of people by a huge bureaucracy, by technocracy
I don’t want the U.S. to compete with China in AI over who can build better tools to censor, repress, and surveil dissidents and minority groups. But that’s exactly what Palantir is trying to do in the U.S. with its efforts to create huge, unaccountable data troves to help policing systems and ICE deportations.
I want us to compete with China in AI over who can build the better industrial Internet of Things and the privacy-preserving algorithms that will help sustain a more trustworthy AI ecosystem.
I don’t have an overly intellectual justification to recommend this week’s must-watch nor is it even loosely related to China’s AI scene, but please go watch The Farewell, directed by Lulu Wang. One of these days mediocre movies about Chinese Americans will get traction but until then we need movies like The Farewell to be extremely excellent in every way and it seems like it even exceeds that high bar.
Using Microsoft’s Building Footprints, the New York Times built a map of every building in the United States.
Great piece by Christopher Magoon on mental health stigma in China and the potential of virtual reality applications to help. My one quibble was with this line “This fact has led some experts to believe China’s AI medical technology will soon surpass the U.S.’s in terms of sophistication and adoption—if it has not already done so” which I just don’t think can be supported — the degree of electronic health records and standardization of health care data in China just cannot compare to the U.S. and other developed countries. I’ve come to question any claim that China is leading in some vertical of AI — very few actually come with comparative metrics that distinguish between different parts of the AI value chain, very few talk about actual commercialization/adoption/diffusion, and almost every one just cherrypicks a couple of anecdotes.
Check out Deeptracelab’s weekly newsletter on the evolution of synthetic media technologies, weaponized misinformation, and new cybersecurity threats. Their most recent issue investigated Chinese deepfake pornography.