Proactive Physical AI Startup, Apple Intelligence in China, and Faster E-Commerce Agent Deployment

AI is shifting from a competition over model capabilities to real-world deployment: on one end is physical AI that can proactively identify needs, and on the other are agents in phones, office work, and e-commerce operations. Apple Intelligence's entry into China has entered its final stage, Jiyuan Ludong and Qianwen Office are accelerating commercialization on developer platforms and enterprise agents respectively, and overseas hardware brands are also using full-funnel storytelling to win global users.^[2][3][6]

Proactive Physical AI: From 'Taking Orders' to 'Finding Needs'

Tang Duyu, former chief AI expert at Huawei and former head of the Xiaoyi LLM, founded Guanxing Shuzhi and proposed 'proactive physical AI.' In the past, physical AI mostly focused on how to execute human instructions, making robots walk more steadily and grasp more accurately; Tang argues that such systems are still 'passive,' solving only how to complete tasks rather than autonomously judging what needs to be done and when to act. Guanxing Shuzhi's technical path has two parts: StreamingJev continuously reads event streams in the real environment, making lightweight judgments like a sentinel about whether intervention is needed; the Response Model acts like a commander, judging the nature of events, planning response methods, and deciding which devices to invoke. Its data approach extracts scene changes from public videos and converts them into decision sequences with reasoning processes, enabling the model to distinguish between 'a change has occurred' and 'a need has formed.'^[1]

The Hard Parts of Proactive AI: Boundaries, Timing, and Dedicated Research

Proactive AI will not simply emerge as large model capabilities improve. Tang argues that it also involves risk assessment, responsibility boundaries, and timing of action, and requires dedicated research to crack. The same action in a physical environment can have completely different meanings: a person suddenly sitting down on the ground might just be resting, or might have fainted; a child reaching toward a cup might just be observing, or might knock it over. A reminder one step early easily becomes an interruption, while one step late loses its meaning. The Guanxing Shuzhi team maintains deep collaboration with a Harbin Institute of Technology laboratory. Tang himself was a PhD student at HIT's SCIR research center, advised by Professor Qin Bing, studying how machines understand human emotion. After receiving his PhD in 2016, he joined Microsoft Research Asia, leading related research on CodeBERT, CodeGPT, and other projects, with cumulative citations exceeding 30,000. Now, facing unconverged routes such as VLA, world models, and WAM, he chooses to bet early on proactive intelligence in the physical world.^[1]

Apple Intelligence in China: Rolling Out with iOS 27.3 in December, with Alibaba and Baidu Involved

Apple Intelligence's launch in mainland China has entered its final stage. According to Bloomberg, Apple plans to officially launch Apple Intelligence through the iOS 27.3 system update this December. The first batch will be AI features already released in the United States two years ago, covering iPhone, iPad, and Mac, but not including the all-new Siri AI assistant released this year. The service passed generative AI service filing with cyberspace authorities on July 8 and obtained launch approval on September 29. On partnerships, Baidu is involved in search and Siri-related functions, and code analysis shows a 'Baidu Visual Search' module in iOS 27 Beta 2; Alibaba will provide Chinese users on iOS, iPadOS, macOS, and visionOS with text and image understanding, content generation, and other experiences, with licensing covering Tongyi Wanxiang and Qwen. The China-market version will most likely require localization from underlying models to content rules.^[2][4]

A $5 Trillion Market Cap and a 'Late' AI Exam

Apple has recently officially crossed the $5 trillion market cap threshold, becoming the world's second publicly listed company after Nvidia to close with a market value above $5 trillion. On that day, US inflation data exceeded expectations, US Treasury yields surged, the Nasdaq fell 1.9%, and tech stocks including Nvidia, Meta, Microsoft, Google, and Oracle fell collectively, while Apple climbed to $347.07 intraday and closed with a market value of $5.005 trillion. Meanwhile, Gurman reported that Apple is preparing to bring Apple Intelligence to mainland China with iOS 27.3 in December. The first batch of features is expected to include Genmoji, custom emoji, notification summaries, writing tools, and photo Clean Up, but the new Siri AI, the redesigned Image Playground, and others are not expected to arrive at the same time. In the Ministry of Industry and Information Technology's network access license information, three AI systems have appeared: Apple Intelligence, Tongyi Qianwen, and Tongyi Wanxiang; as domestic phones have pushed AI toward agents, cross-app operations, and system-level task execution, Apple's $5 trillion exam still has to answer questions about localization and feature pacing.^[4]

Jiyuan Ludong: 300,000 Developers and a Multi-Model Gateway

AI infrastructure company Jiyuan Ludong announced that more than 300,000 developers have joined and launched a limited-time Double 11 presale benefit. Users who make a single recharge of at least 50 yuan on its TokenRhythm API platform receive an equal amount of model invocation credit and can use one API Key to call all models on the platform; during the campaign, users can participate multiple times, cumulative gifted credit has no cap, this round of recharge has no cap, and gifted credit is permanently valid, with the validity period of historical gifted credit also adjusted to permanent. The 'Invite Friends' campaign gives invited friends 18 yuan in credit after they register and complete real-name verification. Jiyuan Ludong positions itself as China's OpenRouter, displaying model configurations, prices, and capability evaluations through a unified gateway, and pairing with its self-developed open-source agent OpenSquilla for multi-model intelligent fusion routing. Its API platform began public beta in August 2026, signed strategic cooperation agreements with Infinigence AI and Alibaba Cloud in September, and recently completed tens of millions of dollars in financing.^[3]

Qianwen Office E-Commerce Edition: Agents Go Deep into Business Operations

Qianwen Office has launched an e-commerce edition, positioned as an independent third-party e-commerce agent, open to all e-commerce platforms and merchants. Its four functions revolve around business operations: trend insight and selling-point analysis find seeding directions from trending terms and popular content on Douyin and Xiaohongshu; batch generation of viral creative assets can incorporate brand colors, visual styles, and selling points into generation requirements; a cross-platform operations dashboard aggregates sales, traffic, inventory, and refunds; and cross-platform ROI intelligent optimization uses scheduled tasks to obtain ad reports and alert on anomalies. Early user Liu Xianglong used Qianwen Office for market research and product detail pages, which used to cost tens of thousands of yuan but now costs only a few hundred yuan in tokens and enables same-day product testing. In one case, 17-person 8Bit Coffee covered more than ten e-commerce platforms and used a product-selection research Skill to find that the sun-dried processing method ranked first, launching a new Geisha cold brew whose sales were about 10 times those of similar past products; a two-person underwear team shortened its autumn/winter asset preparation cycle from 4 months to about 15 days and cut the time for a single data analysis session from about 3 hours to under 1 hour.^[6]

Going Global Hardware: High-Ticket New Products Break Through with Full-Funnel Storytelling

Smart security brand eufy won the 'Best Full-Funnel Marketing Award' at the second YouTube Marketing Awards for its overseas marketing strategy. In early 2025, the North American smart security market faced slowing demand for traditional bullet cameras and intensifying competition. eufy launched two high-end flagships, the eufyCam S4 and PoE NVR S4 Max, focusing on a 'bullet-and-PTZ in one' design that combines a fixed wide-angle bullet camera with an automatically rotating and zooming PTZ camera, priced above $200. After researching more than 3,000 real feedback posts on overseas social media, the team found that users did not reject innovation; rather, they were troubled by how to balance panoramic wide-angle coverage with fixed-point deterrence, and worried about surveillance blind spots, installation difficulty, battery life, and long-term cloud subscription costs. eufy therefore made YouTube the core, advancing through three stages—awareness, consideration, and conversion—and worked with more than 100 creators on real-scenario tests, ultimately capturing high-intent audiences at Black Friday to complete conversions.^[5]

Common Trend: Scenarios, Funnels, and Localization Determine Deployment Speed

From Guanxing Shuzhi's event-stream judgment to Apple's localized AI deployment, and then to Jiyuan Ludong's multi-model routing, Qianwen Office's operations agents, and eufy's full-funnel global push, one common change is that AI value realization increasingly depends on understanding specific scenarios, specific processes, and specific hardware. Going forward, worth watching is how the data system and evaluation standards for proactive physical AI will be established, whether the first batch of China-market Apple Intelligence features can gain market recognition, whether developer platforms and enterprise agents can form sustainable incremental value on top of existing software, and how Chinese hardware brands can convert technical selling points into long-term trust among global users.^{[1][2][3][4][5][6]}