A cofounder explains why Sensetime's strategy differs from that of Megvii and Huawei
|Jeffrey Ding||19 hr ago||3|
Greetings from a world where…
…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). As always, the searchable archive of all past issues is here.
Feature Translation: Sensetime’s Open Source Strategy
This week, we’re taking a look at an August 2020 article by the China Software Developer Network (CSDN), the largest software developer community in China, which also provides IT news coverage.
Context: Longtime readers will recall a white paper on China’s development of AI open source software, which we translated in ChinAI #22. As the white paper emphasized, there is a push for “independent and controllable” (自主可控) AI software, as the main open-source deep-learning frameworks are developed by U.S. companies (Facebook’s PyTorch and Google’s TensorFlow).
At the start of 2020, Megvii and Huawei open-sourced their respective deep learning frameworks: MegEngine and MindSpore. Now everyone’s attention is on Sensetime, the AI unicorn with 4,000 employees. What will it do with its independently developed deep learning platform — SenseParrots? To find out, CSDN interviewed Prof. Dahua Lin, a co-founder of Sensetime.
Key Takeaways: Sensetime’s open source strategy is different from those of Megvii and Huawei. Instead of open-sourcing SenseParrots, its bottom-layer deep learning framework, it has chosen to first open-source its OpenMMLab, which includes upper-level algorithms and application platforms in deep learning for computer vision. The OpenMMLab Github repository has racked up 20k stars on Github (for reference, TensorFlow has 154k, and MegEngine has 3.8k)
Why OpenMMLab and not SenseParrots? PyTorch and TensorFlow have consolidated huge network effects and first-mover advantages at the bottom layer. Thousands of papers and products are built on these two frameworks, and these are rich ecosystems. Simply open-sourcing a training framework is not enough to compete; PyTorch and TensorFlow support multiple layers of algorithms, tools, and various engineering environments. Instead, Sensetime chose to compete at the layer of algorithm and engineering development for building upper-level applications in deep learning for computer vision.
What’s the grand strategy? Dr. Lin claims, “Once we have ecosystem-level influence, we can use this as an entry point to plan the next step. In the future, we will open the lower-level deep learning framework SenseParrots at the right time.” The idea is that once many upper-level applications are built with OpenMMLab, then the migration costs for developers will be greatly reduced once Sensetime open-sources the SenseParrots infrastructure.
Other interesting tidbits: Dr. Lin discusses the main advantages of OpenMMLab, his hopes for something like the generative pretrained transformer (GPT-3) in the computer vision field, and his thoughts on the future division of labor in an AI ecosystem built on large models like GPT-3.
ChinAI Links (Four to Forward)
From the Ranking Digital Rights (RDR) project, Rebecca MacKinnon writes: “Chinese companies have taken meaningful steps over the past five years to protect consumer privacy and security from threats that are unrelated to Chinese government surveillance . . . Their progress shows that Chinese companies can and do respond to pressure from foreign governments, investors, and even users. But the extent to which they can improve policies and practices affecting users’ human rights is severely handicapped by China’s legal and political system. The hard reality is that Chinese companies are powerless to protect users (whether they are in China or abroad) from digital rights violations by one of the most powerful—and unaccountable—governments in the world.
Should-attend: Audrey Tang on Taiwan’s digital democracy, collaborative civic technologies, and beneficial information flows
The next session in GovAI’s webinar series will go live on Tuesday, March 9th (1000-1130 GMT). Audrey Tang, Taiwan’s first Digital Minister will discuss collaborative civic technologies in Taiwan, their potential to improve governance and beneficial information flows, with Hélène Landemore, an Associate Professor at Yale, as a discussant.
Should-read: Citi Can’t Have Its $900 Million Back
An absolutely unreal “horror story about software design.” Citibank accidentally wired $900 million to some hedge funds last August. Some did not send the money back. Citi sued them and recently lost the suit. Matt Levine read the judicial opinion and wrote about the case for Bloomberg. What shocked me was the rigidity of the software that Citi uses to process its financial translations (Oracle’s Flexcube). I think for every article about the newest groundbreaking advance in deep reinforcement learning, we need 10 articles about just basic software development.
China Money Network translates a detailed report by Chinese media 36kr on the scam that was Wuhan Hongxin Semiconductor Manufacturing Corporation . A tale on the perils of China’s industrial strategy on chips — h/t to ChinAI subscriber Ben Mueller for pointing me to the story. For related coverage, see this Caixin article. Careful readers of ChinAI will recall that we covered Hongxin’s troubles last November.
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 PhD candidate in International Relations at the University of Oxford and a researcher at the Center for the Governance of AI at Oxford’s Future of Humanity Institute.
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 email@example.com or on Twitter at @jjding99