ChinAI # 71: What I Got Wrong re: Entity List & Chinese AI Startups

Plus, what does the French Ministry of the Armies Report on AI say about China?

Reflections on the Entity List Ban

In one episode of the Sinica podcast (I forget which one it is now), one of the hosts — I think it was Kaiser — said that studying China is a continual process of saying I don’t know in better ways. That comment was probably from at least five years ago, but it still resonates with me. Using that approach as a guide, I reflect on my comments for the WSJ article I shared last week, and look at what I got wrong and what I can do better in the future.

CONTEXT: A couple of weeks ago, the U.S. added eight Chinese companies — including facial recognition startups Sensetime, Megvii, and Yitu — to a trade blacklist that cuts off their access to American-sourced products and technology without a license. Here’s how the WSJ and I interpreted the impact on these companies:

A major challenge for these companies now will be to secure access to cutting-edge processors, where the market is currently dominated by American companies. That includes powerful graphics-processing units made by Nvidia Corp., of Santa Clara, Calif. The chips are used by AI companies to crunch the massive amounts of data that feed algorithms underpinning AI programs. Nvidia is widely seen as a leader in the field without a clear non-U.S. alternative.

“There’s a reason everybody uses Nvidia GPUs—they’re the best in class,” said Jeffrey Ding, a researcher at the Centre for the Governance of AI at Oxford University. “Even if you switch, it’s going to have a cost on your efficiency; it’s going to have a cost on your margin.”

WHAT I GOT WRONG: I got two things wrong here: 1) Because Nvidia uses TSMC’s Taiwanese fabs to their manufacture chips, firms relying on Nvidia chips will likely not be affected by the ban; 2) AI companies do not necessarily use Nvidia GPUs because they are more efficient; they use them because of architecture lock-in. I thank Doug Fuller and Cooper Pellaton for their criticism on these two points and follow-on comments.

1) The Entity List’s Impact on Nvidia chip sales to Chinese AI firms:

As Doug pointed out in this thread, the hype over the effects of a chip ban on these companies is overblown. My understanding was that even if chips were made at TSMC fabs in Taiwan, they would still fall under the ban if more more than 25% of the content (including IP and software) came from the United States.

Doug replies, “The weird thing is that manufacturing outside of the US automatically makes these manufactured goods fall outside of the 25%. I have to look up the legalese again AND it ain't clear but that is the interpretation both the DOC [Department of Commerce] and its commercial ‘foes’ [U.S. suppliers affected by the ban] have arrived at.” Thanks to Doug for letting me quote him for this post. The legalese here is that “U.S. origin” is based on a 1990s era concept of inputs to physical goods, which does not account for software and IP.

In an internal letter to staff cited in this excellent South China Morning Post article covering the effects of the entity list ban, Megvii’s CEO Yin Qi wrote, “The specific impact is that we can't directly buy products subject to US export regulations, such as x86 servers and GPUs made in the country.” Doug’s interpretation is that since Nvidia’s chips aren’t made in the United States, they do not fall under the ban.

While I think there’s still a lot of uncertainty surrounding what falls under the ban and what doesn’t, I should have made sure I was much more confident about how Nvidia’s GPU sales would be affected before making my comments.

2) Why do firms use Nvidia? We need to get a deeper story than my broad claims of better efficiency. Thanks to Cooper for writing me a 1500-word exchange on this topic and letting me quote him. He’s a student at Georgia Tech who’s worked at Cigna and Alibaba.

The key here is understanding architecture lock-in, specifically CUDA (compute unified device architecture), which is a software layer created by Nvidia that enables machine learning researchers to use a GPU for general purpose processing. Essentially what interface libraries like CUDA do is turn high-level machine learning (ML) operations into low-level instructions that can be run on a card at training time.

According to Cooper, when ML became mainstream, Nvidia’s CUDA interface layer was better at doing this type of interfacing, so ML researchers could turn their workloads into massively parallelizable, runnable jobs with CUDA bindings (using Nvidia’s middleman implementations) and not “worry about micro-architectural changes that come with each year’s revision to physical GPUS.”

Again, the story here is not about efficiency. Cooper writes, “In fact, Intel has surpassed the performance of most GPUs with their Stratix lineup (Stratix 9 was the first to give them a run for their money, Stratix 10 can match). Xilinix has also been pushing hard in this space…NVIDIA is good, but people don’t buy NVIDIA cards because they’re good — they’re big, hot, and power hungry. People buy them because of vendor and library lock-in.”

What Nvidia has done in the past does not guarantee its future. Cooper argues that while lack of access to lots of GPUs might hurt Chinese researchers in the short term, there is now a large effort to make the interfacing easier (porting models to different underlying architectures and libraries). He concludes, “Researchers realize they need to mitigate this dependency on specific-provider GPUs for training and there are a number of efforts to make the transition effort easier. Will these bear fruit in the short term? Likely not. But in 5-10 years as libraries become further abstracted, and workload patterns are better established I would reckon that NVIDIA is fighting a mighty battle against the provider of some XPU with a fundamentally new (surrounding) architecture.”

3) To complicate things even further:

The previous two points illustrate how hard it is to implement economic statecraft in the context of a) complicated, global supply chains, b) uncertainty surrounding the underlying technology. In many cases policies that seem to work or make us feel good on the surface level will ultimately be counterproductive for American interests. As Doug points out, “The export ban actually encourages offshoring as currently written and interpreted.” The entity list actions have hammered home the possible consequences of being dependent on a sole-source supplier for GPUs, and now more firms and researchers around the world (including in the U.S.) are working on alternatives.

Missing among all of this is how the entity list action will actually affect the situation of people suffering in Xinjiang (see the ChinAI links for firsthand accounts). I find it pretty disheartening that so much of this has been framed as just another chip for Trump to play around with in trade deal negotiations, rather than as a coherent, principled formulation of how the U.S. wants to enforce a reasonable standard against the complicity of U.S. companies in human rights abuses.

We’ve only scratched the surface of how complicated this stuff is. I’ve limited the analysis in this reflections on AI software firms (Megvii, Sensetime) access to hardware for training algorithms. We haven’t even talked about hardware firms affected by the ban (Hikvision, Dahua), access to hardware for inference (running AI algorithms), etc. What we can do is continue saying I don’t know in better and better ways.

Feature Translation: China-related Excerpts from the French Ministry of the Armies report on AI

I went through the French Ministry of the Armies (formerly French Ministry of Defence) report on Artificial Intelligence, released last month, and translated the portions that mentioned China. See our GovAI summer fellow Ulrike Franke’s broader analysis of France’s new military AI strategy here:

There’s a strong emphasis on the two AI superpowers narrative. Note as well the emphasis on sovereignty and ensuring France is free from this duopoly. Excerpts from the report:

  • Two "superpowers": the United States and China, out of reach of the others, controlling an immense mass of data, having an ecosystem articulated around powerful integrating companies with global reach (GAFA and BATX) and being able still to multiply their domination thanks to their scientific and financial means;

  • “Outside the European continent, other major states want to free themselves from the duopoly over AI exerted by China and the United States.” (Hors continent européen, d’autres États importants souhaitent s’affranchir du duopole IA exercé par la Chine et les États-Unis)

  • As underlined by the strategic review of defense and national security "... the mastery of artificial intelligence will represent a stake of sovereignty, in an industrial environment characterized by rapid technological innovations and today dominated by foreign companies." The global artificial intelligence ecosystem is dominated by major US and Chinese digital players who are developing internal capabilities and buying up many promising companies.

To be honest, I was underwhelmed by the China-specific portions of the report.

Re: the two AI superpowers claim, I’ve argued elsewhere that this is a) a nonsensical claim without specifying what aspect of the abstract catchall term AI that we are talking about, b) not true based on any reasonable specification of AI, and c) a dangerous meme. In the comments of the full translation, I also point out how there’s not very in-depth analysis or understanding of context of cross-country differences in civil military fusion.

Note: my French background consists of barely surviving AP French, doing a summer stint at the U.S. Embassy in Senegal, and watching a bunch of French-language movies and TV shows that somehow always feature Gerard Depardieu.

FULL TRANSLATION: EXCERPTS OF FRENCH MINISTRY OF THE ARMIES REPORT ON ARTIFICIAL INTELLIGENCE

ChinAI Links (Four to Forward)

Must-read: Weather Reports: Voices from Xinjiang by Ben Mauk for The Believer

“Weather Reports” collects Ben’s interviews of former detainees in Xinjiang’s camps, and the “weather code” they use to check in on how their relatives back in China (if the weather is good so is the health of their relatives).

Context from the piece: These firsthand accounts were recorded in Almaty, Kazakhstan’s largest city, in collaboration with the volunteer-run human rights group Atajurt. In spring 2019, Kazakhstan’s government curtailed Atajurt’s activities as part of a crackdown on public criticism of China…The speakers here have nevertheless chosen to use their own names and the names of their relatives, despite the risks of censure in Kazakhstan and of reprisal for family members still in China. In doing so, they hope to pressure the Chinese government to reconsider its policy of mass detention in Xinjiang.

Must-read: Social Credit Scores in Xiamen and Fuzhou

In the Berkman Klein Center’s collection on Medium, Dev Lewis and a group of Yenching Scholars interviewed officials in Xiamen and Fuzhou, two cities in Fujian which are two of a handful of Chinese cities with their own city-level personal credit scores. h/t to Johanna Castigan for recommending this to me.

Key points:

  1. The data comes from public not private sources - Fuzhou’s platform collects data from over 630 sources, including government ministries, public bodies, state-owned enterprises, but people’s scores are not impacted at all by data from the private sector (online purchases or social media posts).

  2. AI plays no role for now - neither of these models use machine learning based technologies such as predictive scoring, though both governments claim to be bringing in ML. This lines up with Shazeda Ahmed’s comments in a previous issue of ChinAI: “I've been cautious about how I talk about technology in relation to the social credit system because even though I've spotted a few vague, aspirational policy references to ‘implement big data and AI into the credit system,’ thus far the applications I've seen are limited and fragmented.”

  3. A very rudimentary system: Dev states that implementation so far reveals a very basic attempt with numerous gaps and question marks, but a far cry from the western media picture of an all-encompassing score enabled by mass surveillance.”

Go deeper: The post cites Chinese legal scholar Xin Dai’s important paper “The Reputation State: China’s Social Credit Project,” in which he frames the social credit system as an effort of China’s developmental state to tackle its persistent governance problems with reputation-based schemes.

Should-read: MIT Tech Review Interactive and DeepMind Technical Paper on fairness in COMPAS pretrial risk assessment tool

Trained on historical defendant data to find correlations between factors like someone’s age and past criminal record, the COMPAS tool is used in the US court system to help judges determine whether a defendant should be kept in jail or be allowed out while waiting trial. It has come under fire for being biased against African American defendants. The MIT Tech Review interactive, by Karen Hao and Jonathan Stray, nicely illustrates how satisfying different definitions of fairness in this context is impossible.

Silva Chiappa and William S. Isaac point out that debates over the COMPAS tool should also consider patterns of unfairness underlying the training data. In their paper they outline a method (causal Bayesian networks) to measure the unfairness in a dataset.

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 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 & 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