ChinAI #188: Should Chinese Tech Giants Tear Down their AI Research Institutes? (Part II)
Yuchen Li's affirmative case, continued
Greetings from a world where…
agoras aggregate
…As always, the searchable archive of all past issues is here. 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).
Feature Translation: Why I support the BAT tearing down their “AI Research Institutes”
Context: In the previous issue, we started to dig into a fascinating commentary by Yuchen Li, a well-connected Leiphone reporter, that argues for Chinese tech giants to dismantle their AI research institutes. Li’s piece on the AItechtalk platform stated, “The AI research institutes of the Chinese Internet giants have reached a point where ‘there is no construction without destruction’ (不破不立).”
Before continuing with takeaways, I wanted to circle back to something a reader commented on the translation Google doc. A key thread from the first half of the translation was the conflict between independent AI research institutes and the people responsible for implementing AI in various product departments.
Li cited one such dispute between two groups working on voice recognition at Baidu — one led by Andrew Ng at Baidu’s AI lab and the other led by Jia Lei in Baidu’s search division.
One reader emphasized that this Baidu example was an over-simplification. I agree. Care should be taken to understand some of Li’s examples as industry chatter that is meant to illustrate his broader points about organizational structure.
Key Takeaways from the second half:
Recall that constructing independent AI research labs within a functional organizational structure comes with trade-offs: insulation to pursue pure research but isolation from practical, engineering-oriented implementations. To resolve this, some Chinese tech companies have tried to set revenue KPIs for their AI labs:
In 2021, Alibaba set a revenue KPI of up to the 1-billion (RMB) level for its DAMO Academy (once-hyped to take on Google’s AI labs for global leadership). According to Li, this move sparked much internal controversy, leading to the resignation of many high-level scientists.
WeBank, a Tencent-backed bank that has invested a lot in AI, has set up an insourcing system: if a product line has AI needs, its AI research institute will charge the relevant department based on labor and compute costs.
Here’s the rub: “In order to achieve the revenue KPI, the research institute charges indiscriminately for internal projects, despite uneven delivery quality and lack of project experience. On the other hand, the product department is forced to spend a lot of its excess budget because of (data) security considerations. Plus, these projects often fail to meet expectations.”
Li Xiang (pseudonym) told Yuchen that the relationship is imbalanced, since the AI researchers are the “uncles” and can point to their higher rank and salary.
DAMO Academy eliminated its revenue KPI this year.
Let’s say a company is able to implement an effective insourcing system for its independent AI lab. Under the function-based model, this can blind the lab to where the AI frontier is going. Case in point: Siemens Research (USA)
Liu Hua (pseudonym) worked at the Siemens lab for over a decade. In Siemens’ insourcing system, 30 percent of their salary came from headquarters, while “the other 70% of the funds had to be earned by ourselves, and we had to sign contracts with the product department to get projects.”
But Siemens Research (USA) was slow to catch on to the deep learning trend. Here’s Hua again: “We started paying attention to deep learning very early. In 2006, Yang LeCun also visited our laboratory once. In 2012, deep learning succeeded in computer vision (proposed by AlexNet). However, it was not until a 2013 top industry conference, when we discovered that deep learning has invaded our field, that we really started to react and set up a research team of 5-6 people.”
One possible alternative to the functional organizational structure is the matrix structure:
Image below shows that the AI lab reports to both the CTO and the various business departments (first three vertical columns).
Li claims that the matrix system doesn’t work because project scoping gets too complicated, as both the CTO and product division leaders have to sign off on everything. Circling back to the Facebook example, he says that Facebook’s AI lab struggled with serving two masters, and that’s why it ultimately decided to shift toward…
The divisional system, or what Li dubs “the only road left for the AI research institute of Chinese Internet giants”:
Image below displays an AI team assigned to each product line, with no independent AI lab.
In Li’s view, back in 2016, when every Chinese tech giant was racing to scoop up AI talent, the functional system was “obviously correct.” Independent research institutes were needed to attract talents to join.
Now that foundational AI platforms and capabilities have been established, Li thinks “the mission of AI research institutes of major companies has been basically completed.” It’s time to “perform surgery” on AI research institutes and switch to the divisional structure: “Imagine when the AI experts of research institutes are transferred full-time to Taobao, WeChat, Search, Douyin and other products, the former gets the scenario data, and the latter has the top AI algorithms, which will achieve the expected effect of 1+1>2.”
FULL TRANSLATION: Why I support the BAT tearing down their “AI Research Institutes”
ChinAI Links (Four to Forward)
Should-read: Vast Cache of Chinese Police Files Offered for Sale in Alleged Hack
By Karen Hao and Rachel Liang, in The Wall Street Journal:
A vast trove of data on Chinese citizens allegedly siphoned from a police database, some of which checks out as legitimate, is being offered for sale by an anonymous hacker or hacking group. If confirmed, it would mark one of history’s largest leaks of personal data.
The cache allegedly includes billions of records stolen from police in Shanghai, containing data on one billion Chinese citizens, according to a post advertising its availability that was published on Thursday by the hacker on a popular online cybercrime forum. The post, which began circulating on social media over the weekend, put the price for the leak at 10 Bitcoin, or roughly $200,000.
Should-read: Machine Learning Predictions on the 20th Politburo Standing Committee of the Communist Party of China
Lee Johnhyuk, an Assistant Professor at Nanyang Technological University, uses machine learning techniques to make predictions about who is most likely to become a Politburo Standing Committee member. Nifty application of ML in political science context that draws on the CCP Elite database maintained by the 21st Century China Center at UC San Diego.
*H/t to longtime reader Charmaine for this recommendation
Should-read: Panda Paw Dragon Claw Newsletter
Tianjie Ma, who did great work with Chublic Opinion, founded this project: “The Paw Tracker newsletter, developed by Panda Paw Dragon Claw, provides up-to-date and granular project-level information on the Belt and Road Initiative. Drawing from Chinese sources of information that are often disjointed and difficult to access, the newsletter also aims to become a convening space for watchers of the BRI to share and cross-check information about projects and their impacts on the ground.”
Should-apply: DeepMind Research Scientist, Long-term Strategy & Governance
From the advert:
On the Long-term Strategy and Governance Team, we help DeepMind and the world prepare for a world with advanced AI. We map out AI’s potential risks and opportunities from a long-term, global perspective. We work to envision and build recommendations for better governance of AI, identifying actions, norms, and institutional structures that could improve decision-making around advanced AI…Particular areas we are looking for include: global governance of science and powerful technologies; international AI regulations and policy-making; technical AI landscape (see “AI Governance: A Research Agenda”); safety-critical organisations; political economy of large general models and future AI services; macro-strategy.
*Application deadline is 6pm BST on August 1st. H/t to Joslyn Barnhart, a senior research scientist working on this DeepMind team, for sharing this with me.
Thank you for reading and engaging.
These are Jeff Ding's (sometimes) weekly translations of Chinese-language musings on AI and related topics. Jeff is a postdoctoral fellow at Stanford's Center for International Security and Cooperation, sponsored by Stanford's Institute for Human-Centered Artificial Intelligence.
Check out the archive of all past issues here & 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).
Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99