ChinAI #197: Qualcomm and the Boy who Cried AI Wolf
AI software's dream of one-time development, any-site implementation
Greetings from a world where…
translations are living things — shout-out to Tim Fist and Karson Elmgren for adding edits and comments to last week’s feature translation
…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: The last stumbling block in the popularization of end-to-end AI
Context: Earlier this month, the World AI Conference (WAIC) was held in Shanghai. The question hanging over the backdrop: when will AI spread extensively? As AI developer Xiao Wang told Leiphone, “Now hearing that the era of AI is coming, it's like hearing The Boy Who Cried Wolf.” Covering Qualcomm’s efforts in this space, this week’s feature translation digs into the stumbling blocks to the widespread implementation of AI. *Note: I tend to avoid translating PR-heavy articles like this one (flattery of Qualcomm’s capabilities is excessive), but there’s enough nitty-gritty details about actually implementing AI that makes it a worthwhile read.
Key Takeaways: The first stumbling block — the lack of sufficiently powerful AI chips
Last week, we talked about controls on chips used to train AI models (e.g. Nvidia GPUs). Here, we’re discussing chips used for inference, that is, to run AI models on end-side applications.
Smartphones (which began to apply AI for taking pictures and operating intelligent voice assistants) became a competitive area for AI chips. In this domain, we’ve seen impressive improvements in inference speed and energy efficiency over the past few years.
Specifically, the article states: “Looking at the latest ranking of AI-Benchmark, Leiphone found that the top nine mobile phones on the list are all equipped with Qualcomm's first-generation Snapdragon 8 mobile platform.” If you go to the benchmark that Leiphone cites, Mediatek’s Dimensity 9000 system is the only close competitor.
The remaining stumbling block: developing the AI software that can work with these chips in various application scenarios
Megvii’s CTO was also at the WAIC and spoke to this, “A core challenge of AI at this stage is that the fragmentation of application scenarios leads to the diversification of algorithms. Diversification of algorithms, on the one hand, means the need for scale. To produce a large number of algorithms, on the other hand, you need to consider how to produce each algorithm at a low cost.”
The key question: can a facial recognition function originally intended for handheld devices be easily adapted to other end-side applications such as cars, PCs or security cameras? Currently, this is hard because most AI chips use a unique instruction set and architecture.
That’s why Qualcomm was promoting its AI chip software stack, which supports all mainstream AI frameworks and runtimes. An AI runtime is the infrastructure on which AI technology is implemented. A good example is SAP's AI Core, which allows developers to preprocess and train models, deploy models as a web service, differentiate which clients can access which AI assets, and integrate with the cloud.
ChinAI Links (Two to Transfer)
Just a few links this issue, as I was at the American Political Science Association annual conference last week. One of the best panels I attended was “Propaganda and Surveillance: Maintaining Legitimacy and Compliance in China,” part of the Chinese Politics Mini-Conference. Vivian Zhan, a professor at The Chinese University of Hong Kong, presented some research on digital surveillance that startled my thinking.
Their findings, derived from a quasi-natural experiment using two waves of surveillance pilot projects coinciding with large-scale surveys of citizens, suggest “that although newly introduced digital surveillance may generate fear and induce compliance among the public, as authoritarian rulers desire, such effect wanes over time and may even backfire by making citizens less submissive in the long run.” The other three papers were also fantastic. For instance, based on data from millions of newspaper articles, Hannah Waight and Yin Yuan presented work that showed scripted propaganda was becoming more common in China, not just among party newspapers but also commercial newspapers.
It’s exciting to see this group grow! That link will get you to details about our mini-conference from earlier this month, which includes presenters and the names of their papers. We also hosted a social hour at APSA for this research community on the politics of emerging technologies.
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 an Assistant Professor of Political Science at George Washington University.
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 firstname.lastname@example.org or on Twitter at @jjding99