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 please please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay for a subscription will support access for all AND compensation for awesome ChinAI contributors like Ru-Ping from this week). We’re at 64 subscribers — the (very arbitrary) goal is to get to 100 by the end of September.

Feature Translation: Huawei launches AI poet "Yuefu"

Right at the start, let’s get a quick disclaimer out of the way: Yes, this development has “grand-strategic” implications, such as for: A) the rate of cross-border technological diffusion for fundamental models like GPT; B) assessment of the strength of Huawei’s “Noah’s Ark” a sub-lab of the 2012 lab, a turning point in Huawei’s R&D ramp-up, and perhaps foreshadowing of the rough waters coming for the company in years to come.

But at least for this one week, we’re just going to leave it at that. We’re going to enjoy reading some Classical Chinese poetry and allow ourselves to actually marvel at what this AI poet can create. Of course, I care about the grand-strategic/political implications of technology (it’s what I research all day), but at the same time I can’t help but feel that we’ve lost our sense of wonder about the beautiful things that AI can sometimes help bring about. Reading some of these poems by the AI poet/translations by Ru-Ping brought me a sense of wonder that I rarely experience (e.g. when reading Wesley Morris’s movie reviews). I hope it does the same for you.

CONTEXT: Back in June 2019, Huawei Noah’s Ark Lab published a paper on arxiv claiming to be the first to employ OpenAI’s GPT (a pre-trained natural language model) to develop a poetry generation system. Good discussion with Miles and Jack from OpenAI in this thread below, including some clarifications that others had used GPT to do poetry generation:

Jeffrey Ding@jjding99

Huawei Noah's Ark Lab launches "AI poet": writing high quality classical Chinese poetry with a generative pre-trained transformer (GPT) -- "first to employ GPT in developing a poetry generation system" arxiv.org/pdf/1907.00151… Stay tuned for more coverage in next week's #ChinAI

September 10th 2019

9 Retweets29 Likes

Regardless, this is definitely the first to employ GPT in generating classical Chinese poetry, which presents some unique challenges in terms of form, tone, rhyming, and pairing. This is another example of how Natural Language Processing (NLP) varies by the Language. It’s not enough to just get to the level where we’re talking about NLP/computer vision/predictive intelligence instead of the buzzword-catchall of AI; we also have to get to the level where we make distinctions between English NLP and Chinese NLP and NLP for low-resource languages. We’ve covered this NLP landscape in past issues.

This week’s piece, by the media platform QbitAI (which is in the ChinAI spotting guide I shared last week), first takes us through the netizen reaction to playing around with the AI poet. It ranged from the legendary poets would cry after reading this amazing piece to “It’s very neat, but I feel that most of what it is expressing is on the syntax level and not getting to the semantics level. It still likely lacks some soul.” It then gave a range of example types of poetry Huawei’s Yuefu can generate, benchmarking Yuefu against another non-GPT-based system by researchers at Tsinghua.

The Poems and their translations: Very excited to introduce some translations of Huawei Yuefu’s poetry by Ru-Ping Chen. A recent UC Berkeley graduate, her creative writing pieces have been published in The Daily Californian and her idiom translations have been published at TutorMing, a site that enables online Mandarin learning. You can follow her on Twitter @roxychen_56.

After some painstaking work translating ten or so of these poems, here are some of her reflections:

(1) I want to make it clear that my translations of these poems reflect only my personal interpretations. Specifically, I wanted to replicate the artistry typically present within poems written by renowned poets (e.g. Li Bai, Li Qingzhao).

Normally, translations should convey what the poet originally intended to get across to readers. Obviously, because I have no existing means through which I can examine what AI desires to get across to readers, I gave myself some creative freedom in translating these poems.

(2) When I translate poems written by actual poets, I usually reference texts that give in-depth analyses of how the poem is structured and the historical implications of such poems. Because these poems were generated by AI, I had no sources that I could reference.

It is important to keep this point in mind because poetry must be evaluated in the context under which it was created, a context that poems generated by AI lack. Poetry is a written representation of the human experience and unless AI evolves to a point where human feeling is possible, these AI-generated poems inherently lack the depth typically present within poems written by humans.

(3) At best, the AI referenced in this article has learned how to replicate the structure of Chinese poetry and mimic the manners in which poets assemble language to convey a degree of artistry. Scholars more qualified than I to translate these poems would likely agree that these replications bear many flaws that may not be readily perceivable by people who do not frequently read poetry.

Personally, I think meaning can be given to these poems only by humans, or else these poems become mere combinations of flowery words. Additionally, I believe these AI-generated pieces give poetry a chance to evolve because scholars/individuals no longer have to restrict their interpretations to historical/literary contexts.

Jeff jumping back in here: +1 to all of Ru-Ping’s points, and I really want to emphasize the last one — the notion that AI-generated poetry will enable poetry an opportunity to evolve, similar to the way that AlphaGo inspired more people to play Go and completely new tactics.

Can you tell the difference between poetry written by AI vs. poetry composed by legendary Tang Dynasty poets ? Midway through the piece, the author gives the readers a chance to pick out the one poem written by a Tang Dynasty poet (the other three were all written by Huawei’s AI poet):

I couldn’t tell and neither could Ru-Ping after skimming it. If any readers can guess the correct answer, tweet it at me, and I’ll personally ensure you get a special reward of somewhat arbitrary and meaningless Internet points!

Now, let’s read through three pieces of AI-generated poetry together — all translations by Ru-Ping Chen:

First up, I really like this Yuefu-composed Cang Tou Shi (Chinese version of acrostic poems) — every line’s first word or phrase can be linked up into a meaningful phrase/sentence, which is often meant to be hidden. In this case, the first words of each line form the title of the poem: 神经网络 (neural networks)

Neural Networks

Allocating divine status to a soul that has passed—it is natural,

Like the classics that preserve the virtues of ancient wisdom.

The astray scripts of the internet try earnestly to preserve their legacies,

A newfound literary wisdom that shall be passed down for centuries.

Next one is a seven-character quatrain based on the prompt of “Summer”:

Summer Days

Winds rustle the bamboo trees and spread lotus flower scents,

Oh, these April breezes—these ponds accumulate no soot.

Though the idle individual may appreciate exquisite scenery,

She laments the wearisome length of summer nights.

We’ll conclude with a lyrical poem. Manjianghong is a reference to a poem supposedly written by 岳飞, a military general who lived during the Song Dynasty. One way to interpret the poem’s meaning is of a scholarly-like figure taking a walk down memory lane and yearning for an unrequited love:

A River of Blossoms: A Stroll in the Park

An instance of rain precedes the morning sun, 

Matters natural to spring few and far in number, 

Inebriated, I take in the scenery of neighboring gardens.

I reminisce, I muse, 

The happy-go-lucky days of my youth,

Having once tread upon greenery in seasons past.

It is early morning yet the flowers looked washed-out,

The fallen flower petals of a myriad of scents.

I care not and rely upon my drunken state with great relish,

My golden cup—a lotus flower,

I have forgotten all.

An unhinged state of being, 

Restraint all but absent.

A delicate yet charming disposition,

A dainty yet fragile figure.

I care not whether that which decorates my hair has withered and fallen,

Whether my hair becomes dishevelled.

The flowers beyond the curtains adorned with swallows,

The honorable drunken literati too small to deserve honor.

I look upon a flower,

Unwilling to comprehend the poem buried in my heart,

It shall endlessly torment me.

***The full translation features other types of Classical Chinese poetry (pentasyllabic regulated verse, eight-line poem with seven characters to a line, Shuidiao Getou, and couplets, etc. Each type with its own rules regarding the number of words, rhymes, tones, and parallelism) — for more wonderful translations by Ru-Ping, see: FULL TRANSLATION: Huawei launches AI poet "Yuefu"

ChinAI Links (Four to Forward)

“Some people, and I am one, feel that Tang (618–907 CE) poetry is the finest literary art they have ever read. But does one need to learn Chinese in order to have such a view, or can classical Chinese poetry be adequately translated?” What a perfect opening paragraph in this NYRB article by Perry Link on the magic in both the art of the poetry itself as well as the art of translating it.

We’re opening up another round of the Governance of AI Fellowship, a 3-month opportunity to get up to speed with the field and hopefully do some cutting-edge research. Apply before October 13th. Start January or July 2020.

This week’s must-read piece, by Benjamin Heinzerling, argues that NLP’s “Clever Hans” moment has arrived. This is a reference to a real-life horse whose trainer claimed could perform arithmetic and other intellectual tasks, but in reality relied on involuntary cues given by its handler. In the same way, Heinzerling’s piece reminds us that as we train increasingly stronger learners, we need to pay increased attention to their ability of exploiting cues and taking unintended shortcuts.

Call for papers for University of Westminster Press volume on AI For Everyone? “This collection brings together critical debates about Artificial Intelligence (AI) to interrogate how we should understand what constitutes AI, its impact and challenges…” h/t to ChinAI reader Angela Lewis for forwarding this call.

Thank you for reading and engaging.

Shout out to everyone who is commenting on the translations - idea is to build up a community of people interested in this stuff. You can contact me at chinainewsletter@gmail.com or on Twitter at @jjding99

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 please please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay support access for all AND compensation for awesome ChinAI contributors). We’re at 60 subscribers — the (very arbitrary) goal is to get to 100 by the end of September.

***Boston/Cambridge folks: I'll be doing a lunch talk at HLS on Sep 12 on "Law, Technology, and China's AI Dream.” Roll on through if interested - more details here. If you’re a subscriber and around, let me buy you a drink!

One quick ask before we get rolling: I usually send out ChinAI on Sunday night/Monday morning but want to some feedback on what day of the week people prefer. Respond on my Twitter poll if you have a strong preference.

Jeffrey Ding@jjding99

What day do you prefer to receive newsletters? My current approach has been to send out #ChinAI on Sunday night/Monday morning but want to get some feedback.

September 8th 2019

The Mission of ChinAI

Every once in a while I get the hankering to make some cheesy reflection post on the state of ChinAI, so here goes:

The mission of ChinAI is to make itself obsolete. If two years from now, we’re still trying to accomplish the current goal — mitigating the Chinese-English language asymmetry on AI-related issues, checking against misperceptions by providing a more objective and comprehensive understanding of China’s AI development — then ChinAI will have failed.

But what we’re trying to address here requires system-scale solutions not individual-level ones. There’s no turf to defend here: please take this model and run with it, do a ChinAI-mold newsletter for a particular domain of AI or vertical or do a ChinAI2 (trust me when I say there’s plenty of material to cover).

I’m very grateful for all those who have contributed to translation, analysis, and crowd-sourced projects (like our current one on ChinAI Company Profiles). In this spirit, I’d like to:

1. Invite people to guest-write future ChinAI issues. In other words, you’d be compensated for putting together and publishing an entire issue. This isn’t a gatekeeping mission, it’s a mission to open the gates up to fresh, underrepresented voices.

2. Share my ChinAI “Spotting Process” doc. This includes: a short overview of my process for finding candidate writings to translate, a list of sources/platforms I regularly look at, and an initial collection of other great translation platforms I follow. Please add comments/suggestions to the doc.

Feature Translation It's Time to Wake Up from the BAT's Smart City Dream

First, some scene-setting:

• Why are smart cities a big deal? They can help improve the efficiency (e.g. smart traffic management) and security (e.g. smart surveillance) of cities. Is China the world’s leader in smart cities? Nah. The North America region dominates the smart city market in terms of revenue.

• What does China’s smart city market look like? Per an IDC report, total market size = 39b USD by 2023. Most investments are going into three key areas: A) flexible energy management/infrastructure, B) data-driven public security governance, and C) intelligent transportation

• What’s the Chinese government’s stake in this? In 2013, the Chinese government designated 90 cities/provinces as pilot smart city projects. This appears to be in line with the approach of other countries (e.g. the U.S. government's Smart City Challenge and the European Union's Horizon 2020 work program, which both funnel support and funds to a particular set of early-adopter cities).

So what’s this "BAT Smart City Dream” that the author, a writer for the smart industry-focused portal “Intellectual Things” (智东西), is referring to?

• The BAT (Baidu/Alibaba/Tencent) and other Internet giants have all tried to get in at the ground floor of smart city construction through software, hardware, investment, and cooperation channels

• Software = services like Alibaba’s city brain, Tencent’s YouTu Skyeye intelligent transportation platform and Youtu SkyEye intelligent security platform, Baidu’s intelligent traffic light projects

• Hardware + investment/cooperation = some smart camera companies have been acquired by the Internet giants (Youdian Technology and China Transinfo), whereas some have reached strategic cooperative arrangements (The big three of Hikvision, Dahua and Uniview)

• One key area to watch (covered in page 5-6 of the full translation): who has more leverage in these software-hardware alliances? This article identifies the BAT’s cloud computing business as the most important aspect of these deals. Full translation includes a list of these software-hardware alliances

Why does the author argue it’s time to wake up from this dream?

• “Although their CEOs and executives have all personally endorsed their smart city-related business many times on stages…their performance can only be summarized as tepid.”

• While Alibaba secured nearly 20 of those 90 smart city pilot projects — feedback from industry professionals described major issues with over-commitments on compute and higher-than-expected needs in investment.

• First key issue is lack of unified standards, which complicates scaling smart city projects beyond a single application in a single city. Yin Jun, VP of Dahua R&D Center says, “Currently, everyone is in the early stages of development, and each company's solutions and design ideas for user needs and application scenarios all come in different forms…Alibaba may hope to manage everything through the cloud but Dahua and Hikvision may prefer to do coordinated computing through edge + cloud computing. In this context, the design and cost components of the differentiated solutions lead to uneven growth of the profit margins in the project…Customer needs are also diversified, and based on these different requests, it is difficult to achieve a unified design approach in the short term.”

• Second key issue is letdowns in implementation projects. The key government policy (“Guiding Opinions on Promoting the Healthy Development of Smart Cities”) doesn’t provide clear indicators. Accustomed to serving users on fast timelines, Internet Giants are still struggling with the long cycles of projects for business customers which may take a decade of demo -> product testing -> QA -> mass production. One very interesting case is the Microsoft-Wuhan dispute. Wuhan was one of the initial 90 cities, and Microsoft won the Wuhan smart city project with a bid of USD$25 million. But Wuhan claimed that Microsoft under-delivered in terms of products and its Azure public cloud only had a usage rate of 12% — this escalated into a legal dispute.

READ FULL TRANSLATION: It's Time to Wake Up from the BAT's Smart City Dream

ChinAI Links (Four to Forward)

This incredible BBC article by Zhaoyin Feng features my long lost name twin, Jeff Ding, who helps run a Twitter account that translates President Trump’s words into Chinese, along with two other volunteers. They are all Trump supporters and hope to “spread Trump’s messages in the Chinese-speaking world.” H/t to Rebecca Kagan at Georgetown’s Center for Security and Emerging Technology. Stay tuned for more cool stuff coming from Rebecca.

NYT Chinese has a really cool daily newsletter, which provides a roundup (in Mandarin) of some of its top stories.

MacroPolo had another praise-worthy week.

• First, this week’s must-read is Neil Thomas’s debunking of the "engagement with China failed" DC echo-chamber narrative. Hey here’s a crazy thought: Instead of pushing the flimsy theory that “younger China hands tend to be more hawkish” without any empirical evidence or comparative stats — as Politico's foreign affairs reporter Nahal Toosi and Bill Bishop have done — Maybe just maybe young China hands don’t subscribe to the simplistic hawk/dove dichotomy and write more clear-eyed takes because they aren't jaded by years of cynicism/constant need to say "I told you so"? Just a thought.

• Second, MacroPolo published an excerpt “2014: The Year of China’s Digital Cambrian Explosion” from Matt Sheehan’s book The Transpacific Experiment. I’m collecting some notes on Matt’s book in this Twitter thread as I go through it, and it’s been a sharp yet sensitive (in all the good senses of that term) read so far.

Really insightful piece on the differences and similarities of Chinese and American military-civil fusion by Elsa Kania for The Strategy Bridge. One snippet: “U.S. defense experts may be surprised that certain U.S. policies and practices are routinely characterized by Chinese military researchers as involving American military-civil fusion, just as Chinese colleagues often claim to be confused why Washington is concerned that China is pursuing a strategy that is seen in their eyes as reminiscent of American approaches.” Also check out her and Helen Toner talking about “Beyond the Arms Race Narrative": AI & China” for the FLI podcast.

Thank you for reading and engaging.

Shout out to everyone who is commenting on the translations - idea is to build up a community of people interested in this stuff. You can contact me at chinainewsletter@gmail.com or on Twitter at @jjding99

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 please please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay support access for all AND compensation for awesome ChinAI contributors). We’re at 60 subscribers — the (very arbitrary) goal is to get to 100 by the end of September.

***Boston/Cambridge folks: I'll be doing a lunch talk at HLS on Sep 12 on "Law, Technology, and China's AI Dream.” Roll on through if interested - more details here. I’ll be sticking around for a couple days and have a good bit of free time on the 12th, so if you’re a subscriber, let me buy you a drink!

Feature Translation (1/2): Why is it that the stuff you search for on Jingdong will appear in ads on Douyin?

This week’s first feature translation is an informative fun explainer about the interaction between two major tech platforms: 1. JD - the Chinese e-commerce giant, 2. Douyin - social media app for short lip-sync, comedy, and talent videos (TikTok is American version). The author, “Chief of Poor Reviews/Chief Chaping” (差评君) is a well-known blogger who writes for Chaping (差评) - new media portal centered around young people’s demand for information on new technologies.

Big ups to first-time ChinAI contributor Justin Tu for doing the bulk of translating on this one. He’s a Chinese Linguist at AECOM and previously interned at the U.S. State Department while he was at Illinois (a.k.a UIUC). He’s also contributing an article on recent Chinese conceptions of AI Safety for national and political security in the upcoming DigiChina Special Report, and is looking for collaborators on the topic!

Chief Chaping recounts how one day he was looking to buy a fridge on JD.com but was interrupted by a work obligation. Later that night, he was browsing Douyin before bed and kept swiping through fridge ads one after another. At first he gets concerned about JD and Douyin stealing/selling his personal data, then after doing some research on ad delivery mechanisms, he concludes “Things probably aren’t as bad as everyone thinks.’

Here’s how it works: When you open Douyin, an ad slot with your unique user label goes live. Per Douyin’s Privacy Policy, it collects:

1. Your actions and behavioral information, such as, Follows, Saves, Searches, Browsing Preferences (e.g., audio and video information of your interest);

2. Your Feedback, Posts, Likes, Comments and other information actively provided by you;

3. Your geographic location information with your express consent.

These user profiles go to a platform (which can be owned by Douyin itself or a third-party) akin to an ad exchange for real-time auctions.

Meanwhile, businesses with demands like Jingdong, Suning [another large Chinese e-commerce retailer], and mobile game titles are the bidders!

The bidders will first attempt matching the user profile in the ad slot with their own user profiles from their database to determine the value of the ad slot. Bidders will generally attempt high bids to snatch a given slot if the profile is an extremely good match.

The above process is called real-time bidding, which is more accurate than the other main type of ad delivery method — labelled targeting, which works through “middlemen” who match ads with ad slots, and is less expensive but also less precise. In my view the key part of this translation is a section on whether user profiles are truly anonymized.

Chief Chaping writes, “Most apps nowadays request IMEI (International Mobile Equipment Identity) from users when they sign up. These IMEI numbers, in turn, act like online IDs for users. Alongside accurate user behavior and interest profiles, platforms can pretty much deliver ads with great precision if they authenticate users through IMEI with each other. Even so, there is no need for panic. This doesn’t necessarily mean that user data is maliciously leaked. As far as the platforms are concerned, they don’t know ‘you’re you.’ You are merely a string of hashed codes.”

One key point of clash: Singapore's data protection watchdog, for instance, issued draft advisory guidelines for telecomms businesses that stated they do not have to treat IMEI numbers as personal data, where those numbers are "viewed in isolation" but they may qualify as personal data where they can be linked to other identifying details of individuals. This is an ongoing debate in China as well. See this past ChinAI translation in which Wang Fang, a data protection officer for Huawei, argues that mobile device identifiers like IMEI numbers do constitute personal data.

FULL TRANSLATION: Why is it that the stuff you search for on Jingdong will appear in ads on Douyin?

Feature Translation (2/2): ChinAI Company Profiles Continues with Inceptio, G7 Networks, and SquirrelAI

In last week’s issue we introduced a new ongoing series of ChinAI Company Profiles; we’ve crowdsourced a Google spreadsheet on these companies here (experimenting with giving edit access to everyone with a link so please don’t troll). This week, we continue three more companies.

The first two — Inceptio and G7 Networks — are linked in an interesting way. Targeting freight transport, Inceptio (嬴彻科技) is an autonomous driving startup focused on L1 and L1.5 trucks (somewhere between a “hands on" systems like adaptive cruise control and a "hands off" systems where the automated system takes full control of the vehicle but driver must be ready to intervene). G7 Networks, the largest shareholder of Inceptio, claims to be China's leading Internet of Things technology company. Official statements say that it has data flows of nearly 1 million vehicles. In the "capital winter" of last year, G7 raised $320 million. Also, Ma Zheren, CEO of Inceptio, is also the president of G7 Networks.

The QbitAI profile of Inceptio emphasizes how this model has attracted many talents including Qi Zichao who graduated from Tsinghua’s renowned “Yao Class” which basically only accepts the “Number One Scholars” from each province and gold medal winners in competitions like the International Olympiad in Informatics (which Qi won in 2009). As I mentioned in my Twitter thread on Megvii’s recent IPO, all its cofounders attended this class, led by Professor Chi-Chih Yao, a winner of the Turing Award who gave up US nationality to return to China.

Other Big-Picture Takeaways:

• Again, the information arbitrage is in the unsexy areas. Not DECOUPLING! but machine visual quality inspection of knife production lines. Not consumer cars but freight transport.

• We are seeing these Chinafornia-model startups (to steal a phrase from Matt Sheehan) pop up more and more, in which a good portion of the R&D research takes place in Silicon Valley (e.g. Qi Zichao is Inceptio’s planning tech lead at its Silicon Valley R&D Center) but the target market is China. I wrote about this in my MIT Tech Review profile of another Chinafornia-style autonomous driving startup, Roadstar.

• This type of model isn’t new (Baidu does a good chunk of their autonomous driving research in Silicon Valley as well) and is not going away anytime soon regardless of how many times pundits scream DECOUPLING! into the void (the U.S. also has overseas R&D labs in other countries as well. Please search up “asset-augmenting R&D” in Google Scholar if you’re really interested in how innovation is globalizing.

• From a U.S. competitiveness perspective, the main concern should be how do startups like U.S.-based startups like Zoox retain or continue to attract the next Qi Zichaos of the world. Some ideas off the top of the dome: appropriately value your bilingual staff as they have twice as many options of places and companies to work for. Upgrade your understanding of Chinese competitors by mandating all your employees read and subscribe to ChinAI?

FULL TRANSLATION: Inceptio’s Past Year: “Yao Class” genius joins, a transport model emerges, and there is already commercialized income

ChinAI Links (Four to Forward)

We’ve also added a working ChinAI Company Profile of Squirrel AI which at this point mostly consists of excerpts from Karen Hao’s must-read MIT Tech Review article on this startup which provides extracurricular tutoring programs with lessons curated by an AI algorithm. Here’s just one of many juicy tidbits: “In the five years since it was founded, the company has opened 2,000 learning centers in 200 cities and registered over a million students—equal to New York City’s entire public school system.”

In a piece for War on the Rocks, My GovAI colleagues Jade Leung, Sophie Charlotte-Fischer, and Allan Dafoe argue that proposed export controls on artificial intelligence are not only ineffective but could also inadvertently strengthen the technology base of U.S. competitors while weakening its own. “Washington appears to be defaulting to traditional, 20th-century policy tools to address a 21st-century problem.” Relatedly, Scott Moore’s Lawfare article also warns that Trump’s techno-nationalism will do lasting harm to U.S. economic competitiveness and innovation.

I’m a little late to this but earlier this month Vox’s Kelsey Piper published a balanced, well-researched (25 hyperlinks to sources!) deconstruction of Peter Thiel's NYT hitpiece on Google’s overseas AI labs in China. She draws from a previous ChinAI issue that lists all the things wrong with Thiels’s NYT oped (it’s a lot of things).

I recently came across Liz Carter’s “Today’s Chinese” project in which she teaches a (usually Mandarin) Chinese word or phrase every day on her Twitter feed. She’s been doing this for about seven years now and is currently getting her PhD in Chinese linguistics at UCLA.

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 please please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay support access for all AND compensation for awesome ChinAI contributors). We’re at 50+ subscribers — if we get to 100 by the end of next month, that wouldn’t suck.

Feature Translation (1/2): Leiphone Best Growth List of AI Companies

In the course of walking through Leiphone’s S&T media empire in a recent issue of ChinAI, I came across a cool ranking called the “AI Best Growth Companies List.” In 2018 and 2019 they awarded three types of best growth awards across ten fields: 1) Product Growth Award, 2) Commercial Growth Award, 3) Future Growth Award. *Note: in 2019 they added the two domains of edge computing and smart cities, and a fourth award called “Best Fortified Growth” (最佳壁垒成长奖). I compiled the two years of award winners into the table below, which gives a cross-section of the rising AI companies in China. The full Google doc translation (also always linked in the section heading) outlines the selection process, criteria, and I make some notes on particularly interesting companies. How many of these companies do you recognize?

Call for Action: I’ve listed all of the above companies along with the unicorns from last week’s issue in this editable Google spreadsheet (anyone with this link can edit so I’m trusting that our readership has no trolls). I’d like to crowdsource help from the ChinAI community in filling out interesting info about each of these companies. You can also add variables of interest as extra columns. Another way you can contribute is to find/help translate good articles on these companies. One of the goals is to have at least one ChinAI issue that features/discusses each company (currently we’ve covered 8 of them so far, including the following translation on Ultrapower (神州泰岳), which won the AI Best Future Growth Award in the AI + Application Platform vertical in both years.

Last time I floated out a call like this, Karson Elmgren stepped up to the plate, and his work was featured in this MIT Tech Review article by Karen Hao. He’s now doing cool stuff at OpenAI. Excited to see who emerges this time.

Feature Translation (2/2): Ultrapower ChinAI Company Profile

This article, also from Leiphone, looks at the 20+ year history of Ultrapower. Founded in 1998!, Ultrapower Software (神州泰岳) rose to fame as the sole software provider for Fetion (飞信), China Mobile’s instant messaging client, which the OGs will recall. Fetion had many challenges — including the rise of this app that you may have heard of (WeChat) — but Ultrapower seems to have pivoted successfully to AI, specifically in the subfield of natural language processing. Some key details:

1) Has open-sourced bunch of Chinese NLP resources, including evaluation datasets and methods

2) Products include a Ruida Control SaaS platform (a financial risk control system for financial institutions), a Taiyue semantic factory to support the internal NLP technical capabilities of companies

3) “AI business is particularly eye-catching in the public security industry”: article lists 6 cases of strategic cooperation between Ultrapower and divisions of public security bureaus at the province and city levels, including efforts to combat “Internet Crimes” in Guangxi Zhuang Autonomous Region and provision of “intelligent reference points” for public security research in Sichuan Garzê Tibetan Autonomous Prefecture. Most coverage has focused on AI-enabled surveillance in Xinjiang (and rightly so), but it’s important to remember that China has 4 other autonomous regions/30 autonomous prefectures (where Ultrapower is doing its business).

4) A lot of Ultrapower’s core business is still in its traditional ICT operations management — this is the not-sexy but super important stuff like making sure everyone’s in your business is up-to-date on hardware and software changes. They’re supposedly making the leap in this field from automation to intelligentization.

Full translation includes some tidbits from a brief scan of other articles featuring Ultrapower — FULL TRANSLATION: Ultrapower ChinAI Company Profile

ChinAI Links (Four to Forward)

This week’s must-read is Larissa Schiavo’s piece on “Lessons in Technological Worker Displacement” which examines the effects of two significant innovations (the power loom and the threshing machine) on laborers in Victorian Era England — with an eye toward parallels to present-day labor-displacing AI technologies. While her piece is really good on technological displacement (underwork), I also want to highlight how these automation-related advances may also lead to overwork — for example, consider the “sweating” system (sweatshops) - labor exploitation (long hours, low wages, unsafe conditions) that emerged in the mass production methods pioneered by the industrial revolution. Raphael Samuel discusses “sweating” in giving the view form labor on the “heroic age of invention.” This reminds me of this recent NYT piece by Cade Metz on the labor-intensive processes that go into labeling data for AI algorithms as well as ChinAI #41 on the data labeling industry in China.

Very cool special issue of Journal of Strategic Studies on emerging technologies and strategic stability, featuring my GovAI colleagues Ben Garfinkel and Allan Dafoe’s really nifty piece, “How does the offense-defense balance scale,” which examines scaling effects (how increases in investments will favor the offense at low levels and favor the defense at high levels) on security implications of AI applications. Are there any services out there that translate academic international relations papers like this collection into digestible bite-sized summaries/takeaways? Would love to get looped into any related projects.

I want to plug a few great reports authored by John VerWey at the U.S. International Trade Commission: a report on the past and present of Chinese Semiconductor Industrial Policy as well as one on the industry’s prospects for future success, and finally one on the potential impacts of Made in China 2025 on the chip industries in the US/EU/Japan (with Dan Kim)

Incredible NYT longform journalism by Amanda Taub and Max Fisher on how Youtube radicalized Brazil — contextualizes well-known effects (e.g. ecochambers) as well as less-discussed ones (e.g. how videos pumping conspiracy theories were recommended to viewers watching videos about right-wing politicians). I’ve started subscribing to Matt Stoller’s “Big” newsletter on the politics of monopoly, and his backstory on this article is worth a read. As I’ve pointed out in previous issues, I’m worried that newsletters can also function as echo chambers, and I’m still exploring ways to ensure more voices are featured in the curating process of ChinAI.

Thank you for reading and engaging.

Shout out to everyone who is commenting on the translations - idea is to build up a community of people interested in this stuff. You can contact me at jeffrey.ding@magd.ox.ac.uk or on Twitter at @jjding99

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 please please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay support access for all AND compensation for awesome ChinAI contributors and collaborators like Joy from this week and others like Charles and Lorand from past weeks.

Feature Translation: CAICT Report on the Global AI Industry

CAICT is a research institute under the Ministry of Industry and Information Technology and one of the co-authors (alongside Tencent Research Institute) of the 500-page book on AI strategy that first launched this newsletter. My collaborator this week is Joy Dantong Ma of MacroPolo. Joy found this report and did the bulk of translating, including some of the key graphs. Her analysis: This report dissects the AI industry into four aspects: company, capital, academic papers, and conferences. It then assesses all major stakeholders, including both institutions and countries, across these aspects. What I find most fascinating is the depth and timeliness of understanding CAICT has on the global landscape. A case in point: in the company section, the report listed 17 unicorns in China, the majority of which are seldom talked about even though China+AI has become such a hot topic. The report also listed out unicorns in the US - Avant, Uptake, Dataminr - that many of us in the US might have never heard of. 

*Also, highly relevant is a project on Chinese AI companies that Joy and I and Matt Sheehan worked on back in December 2018, which goes beyond the abstract catchall of AI and drills down into specific verticals (e.g. autonomous vehicles, voice & speech recognition, business intelligence, etc.)

Anyways, back to the report’s key findings:

 1. As of the end of March 2019, there were 5,386 active artificial intelligence (AI) companies in the world. The US, China, the United Kingdom, Canada, and India rank as the top 5 globally in terms of the amount of AI companies.

2. There are 41 AI unicorns globally, including 17 in China, 18 in the US, 3 in Japan, and 1 each in India, Germany and Israel.

3. Since Q2 2018, global AI investment has gradually declined. The total amount of global investment in AI in Q1 2019 was US$12.6 billion — down 7.3% from the previous quarter, and flat year-on-year. China's AI financing totaled US$3 billion, 55.8% down year-on-year, accounting for 23.5% of total global financing, down 29% from the same period in 2018.

4. Statistics on AI academic papers in the past 10 years: China ranks first in terms of the total number of papers published, while the number of highly cited papers is lower than that in the US.

• Chinese research institutes such as the Chinese Academy of Sciences and Tsinghua University are among the upper echelon of AI academic research institutions.

• Google and Microsoft published the most amount of papers in top AI conferences globally.

FULL TRANSLATION: Global Artificial Intelligence Industry Data Report (April 2019)

ChinAI Links (Four to Forward)

This week’s must-read is a report by Dongwoo Kim (research fellow at Asia Pacific Foundation of Canada) comparing AI policies across China, Japan, and Korea — with an eye toward Canada’s interests. The report emphasizes that Japan and Korea are reliable partners for cooperation in the space of AI (5th and 7th largest trading partner), and that Canada could help bridge the gap between China (2nd largest trading partner) and the West. Also, some really good stuff on Japan’s Society 5.0 and its Strategic Council for AI Technology’s policies as well as Korea’s 30-year “Master Plan” for an intelligent information society.

Had a great time talking about AI race rhetoric, Jessica Newman’s excellent China AI Policy primer, Peter Thiel, relative/absolute gains with Lucas Perry on the Future of Life Institute’s AI Alignment Podcast — super impressed by how FLI produces their podcasts — they have a transcript of the entire podcast, detailed time stamps, and long block quotes as key points. Reminds me of a16z’s podcast about podcasting where they discuss how to improve tools for engaging with podcasts. Jade Leung, my boss and the person who makes GovAI run, was on the AI Alignment Podcast last month to discuss GovAI’s research agenda and what ideal governance in this space looks like.

Based on a public records request to HK’s Government Logistics Department which revealed tenders for facial recognition software, this is really excellent reporting by Rosalind Adams of Buzzfeed on how facial recognition is actually being used by HK authorities: 1) it’s likely that no gov depts have used or tested automated facial recognition as part of its CCTV systems, 2) according to the Immigration Department its facial data has not been shared with the Hong Kong Policy Force. HK has contracted with French company Idemia for facial recognition technology to process Hong Kong ID cards (US State Department works w/ same company on same process). However, while automated facial recognition isn’t being deployed through CCTV, faces are being weaponized amidst the protests, as Paul Mozur reports in this NYT piece.

Rather than centralizing project selection which is what initiatives like the Joint Artificial Intelligence Center do, Eric Lofgren argues we should decentralize the Pentagon’s budget by mission type to ensure AI projects “receive funding at the speed of relevance.” His framework is an important one to consider: “Military capabilities may never benefit from a single general AI application. Instead, they benefit from a variety of narrow AI applications. It seems that the effort spent developing an app for autonomous flight does not contribute much to an app for ground vehicles, let alone automating logistics, target recognition, command and control, or any number of other applications. Each app requires its own data inputs, metric selection, and training.” This was published in War on the Rocks as a response to Eric Schmidt and Robert Work’s call for ideas for the Nat Sec Commission on AI.

Thank you for reading and engaging.

Shout out to everyone who is commenting on the translations - idea is to build up a community of people interested in this stuff. You can contact me at jeffrey.ding@magd.ox.ac.uk or on Twitter at @jjding99