Daily Briefing

China AI & robotics developments, rewritten weekly-context-free for US readers. Published daily at 07:00 Beijing time.

Fri, Nov 6, 2026

Open-weight Chinese AI models near top closed models at hacking, research finds

Open-weight artificial intelligence models are now almost as capable of breaking into computer networks as the best closed proprietary systems, according to new research reported by Wired, a finding that could reshape how US policymakers and security teams weigh the risks of freely downloadable models.

The research compares open-weight models, whose parameters can be downloaded and run by anyone, against closed models accessible only through paid APIs. Wired reported the gap in offensive cyber capability has narrowed sharply. The finding matters for Washington, where export controls and open-source policy are under debate. Chinese developers have released a steady stream of open-weight models that US firms and researchers can download freely. The item did not name the models tested, the researchers behind the work, or the specific benchmarks used.
Source: wired.com · via Physix Frontier briefing desk · Editorial briefing card →
Hundreds of exposed Artex AI hacking tools found open on the web

Security researchers say hundreds of copies of Artex, a Chinese-made AI hacking tool, have been left publicly accessible on the internet, potentially letting anyone use them. The finding highlights how quickly offensive AI tooling is spreading beyond controlled research settings.

TechCrunch reported the exposure on Nov. 6, citing security researchers. The report did not name the researchers or the vendor behind Artex, and did not specify how many instances were confirmed or what systems they could reach. The tools were described as publicly exposed, meaning they could be accessed without authentication. The report did not say whether any had been used maliciously. AI-assisted attack tooling has drawn growing scrutiny from US security teams as lower-cost, automated hacking products circulate online.
Source: techcrunch.com · via Physix Frontier briefing desk · Editorial briefing card →

Thu, Nov 5, 2026

China cross-border logistics firm EDA slides to loss as platforms tighten control

EDA Group's revenue rose 14.3% in the first half of 2026 while its net result swung from a 19.3 million yuan profit to a 42.16 million yuan loss (roughly $2.7 million), according to a Huxiu report, as platforms reclaim last-mile pricing power from overseas warehouses.

The report frames the swing as an industry-wide shift rather than a single company's problem. After platforms moved to semi-managed models, overseas warehouses are left with revenue that increasingly resembles a per-item handling fee, with unit prices set by the platforms. EDA Group is a Chinese cross-border e-commerce logistics provider operating overseas warehouses for merchants selling abroad. The article does not break out segment figures or name the platforms involved. The piece attributes the margin pressure to platforms taking back control of last-mile delivery pricing, leaving warehouse operators with limited room to set rates.
Source: huxiu.com · via Physix Frontier briefing desk · Editorial briefing card →
Perplexity debuts always-on "Teammates" AI agents for workplace tasks

Perplexity has launched "Teammates," always-on AI agents designed to handle multiple work roles, according to The Verge, putting the startup in direct competition with OpenAI and Anthropic in the enterprise agent market.

The agents are described as able to take on multiple roles, though the report did not detail pricing, availability or specific capabilities. The move places Perplexity, a search-focused AI startup, against larger rivals OpenAI and Anthropic, both of which have pushed agent products for workplace use. Details on launch timing and enterprise partners were not included in the report.
Source: theverge.com · via Physix Frontier briefing desk · Editorial briefing card →
Chinese EV maker Deepal patents rear-seat foot bath

Deepal, the electric-vehicle brand backed by state-owned automaker Changan, has filed a patent for a rear-seat foot bath system, per InsideEVs, extending a Chinese habit of packing cabins with feature gimmicks to stand out in a crowded market.

Deepal is a Changan Automobile brand launched in 2022, sold in China and expanding into Europe and Southeast Asia. It is not sold in the US. The patent filing, reported by InsideEVs, describes a foot bath integrated into the rear cabin. Details on production plans were not given. Chinese EV makers have leaned on novelty cabin features, from karaoke to in-car drones, to differentiate in a price war. US readers should treat this as a design signal, not a product launch.
Source: insideevs.com · via Physix Frontier briefing desk · Editorial briefing card →
Tencent releases Hy4 open-source LLM, claiming performance leap

Tencent said it released Tencent Hy4, its latest-generation large language model, which the company describes as a significant jump in performance and capability and one of the leading open-source AI models.

The Shenzhen-based company, known in the US mainly as the WeChat and gaming giant, has been building out its AI stack alongside domestic rivals including Alibaba, Baidu and DeepSeek. Its prior Hunyuan model family has been distributed under open-source licenses. Tencent framed Hy4 as joining "the world's leading open-source AI models," according to its announcement. The release did not disclose parameter counts, benchmark figures or license terms in the material available. Open-weight Chinese models have become a common starting point for US developers seeking lower-cost alternatives to proprietary APIs.
Source: tencent.com · via Physix Frontier briefing desk · Editorial briefing card →
Z.ai says GLM 5.5 trained on 1 trillion web pages

Z.ai's chief scientist Jie Tang disclosed that the company's GLM 5.5 model was trained on roughly 1 trillion web pages, a scale claim that puts the Chinese developer in the same data-regime conversation as leading US frontier labs.

Z.ai is the international brand of Zhipu AI, a Beijing company spun out of Tsinghua University's AI research and among China's earliest large-model developers. Jie Tang, a Tsinghua professor, is its chief scientist. The disclosure came in a post on X, according to the shared link, and offered no further detail on token counts, data mix or compute. Zhipu has previously released its GLM series as open-weight models, a strategy that has driven adoption among developers outside China. Treat the figure as a company claim: web-page counts are not directly comparable across labs, and no independent verification was provided.
Source: twitter.com · via Physix Frontier briefing desk · Editorial briefing card →
20 Chinese tech firms sign AI agent self-discipline pact at Wuzhen

At the 2026 World Internet Conference Wuzhen summit on Nov. 4, the China Academy of Information and Communications Technology and 20 companies including Huawei, Tencent, Alibaba, Ant Group, Douyin and Xiaomi released a self-discipline convention on AI agent development, according to Huxiu.

The signatories span cloud, payments, social and hardware players. CAICT is a government-affiliated research institute that often drafts industry standards before formal regulation. The article says AI agents that operate apps still face three hurdles but does not detail them in the summary provided. Beijing has moved to set norms for agentic AI as such tools gain traction; the pact is voluntary rather than binding.
Source: huxiu.com · via Physix Frontier briefing desk · Editorial briefing card →

Wed, Nov 4, 2026

OpenAI says its security AI found 948 Linux kernel privilege-escalation flaws

OpenAI said on Nov. 4 that a new cybersecurity model, paired with an automated pipeline for finding, validating, exploiting and patching bugs, surfaced 2,308 potential Linux kernel issues and reproduced 1,724 of them. The company reported 948 privilege-escalation vulnerabilities, a result that, if confirmed, would mark a significant step for AI-driven vulnerability research.

The system was described in a security research report published Nov. 4. According to OpenAI, the pipeline identifies candidate flaws, dynamically reproduces them, builds exploits and generates patches. In the Linux kernel test, 2,308 potential security problems were flagged and 1,724 were reproduced, with 948 classed as privilege-escalation vulnerabilities, the company said. OpenAI did not say in the cited summary how many of the findings were confirmed by kernel maintainers or how the patches were received upstream. The Linux kernel is the core of the open-source operating system used across servers and Android devices.
Source: ithome.com · via Physix Frontier briefing desk · Editorial briefing card →
Xiaomi humanoid robots complete 8-hour shift at EV plant

Xiaomi released footage showing its humanoid robots working a full eight-hour shift sorting center console side covers at its electric-vehicle factory, the company said, a demonstration aimed at showing autonomous, sustained factory work rather than staged tasks.

Xiaomi is the Beijing-based smartphone and consumer-electronics maker that entered the EV market with its SU7 sedan. It has been developing humanoid robots alongside its car business, part of a broader push by Chinese manufacturers to deploy embodied AI on production lines. The company said the robots completed the sorting work on their own, according to footage it published. Details on error rates, supervision and the number of units deployed were not disclosed in the source report. Xiaomi joins a growing field of Chinese firms testing humanoids in factories, including automakers and robotics startups, as US and Chinese companies compete over real-world deployment.
Source: electrek.co · via Physix Frontier briefing desk · Editorial briefing card →
China's first fully AI-generated film flops at box office

China's first fully AI-generated feature film earned just 1.7 million yuan (roughly $240,000) at the box office despite costing tens of millions to produce, according to a Guardian report, a result that tests early claims about generative AI's near-term role in commercial filmmaking.

The film's budget was described as "tens of millions" of yuan, far exceeding its reported theatrical return. The Guardian report did not name the production company behind the project. Despite the weak result, some in the industry still view fully AI-generated filmmaking as a viable direction, according to the report, which frames the release as an early commercial test of the format. The outcome lands as Chinese and US studios alike experiment with generative video tools in production pipelines.
Source: theguardian.com · via Physix Frontier briefing desk · Editorial briefing card →
Baidu robotaxi unit touts world model for real-time driving simulation

Baidu's autonomous-driving arm says it has built a world model that simulates whole traffic scenes in real time, including pedestrians with their own behavior, claiming one simulated day equals three years of road testing. The claim, reported by Chinese tech outlet QbitAI, is the latest bid by a Chinese robotaxi developer to cut reliance on costly real-world mileage.

The report frames the scale of the problem in distance: 1 million kilometers, about 2.6 times the Earth-Moon gap, is what an L4 system would need to cover in real road testing. Baidu has operated its Apollo Go robotaxi service across multiple Chinese cities. The item gives no technical details, no named executives and no independent verification of the "one day equals three years" comparison. Baidu did not immediately confirm the figures in the report. World models, which generate synthetic driving scenarios for training and validation, are an active area for autonomous-vehicle developers globally.
Source: qbitai.com · via Physix Frontier briefing desk · Editorial briefing card →
Ex-Tencent researcher's tactile-sensor startup raises tens of millions of yuan

Linggan Robotics, a Chinese startup building vision-and-touch sensors for physical AI, has raised a tens-of-millions-of-yuan angel round led by Leaguer Capital and Hefei Innovation Investment, according to Leiphone. The company says its Lux-series tactile sensors are already shipping at scale.

The founder is a former senior researcher at Tencent, Leiphone reported, without naming them. The round was advised by Zhizhen Consulting. The exact amount was not disclosed; tens of millions of yuan is roughly $1.4 million to $7 million. Tactile sensing is a focus for robot-hand and manipulation startups in China and the US, where touch data is seen as a gap versus vision. Linggan claims volume delivery of its Lux sensors, a step beyond lab demos, though no customer names or shipment figures were given. Investors include Leaguer Capital and Hefei Innovation Investment, per the report.
Source: leiphone.com · via Physix Frontier briefing desk · Editorial briefing card →
Huawei's Xiaoyi Work assistant adopts openPangu-2.1-Pro model

Huawei's Xiaoyi Work assistant has become the first product to run the company's openPangu-2.1-Pro model, according to IT Home, positioning the HarmonyOS task assistant around Huawei's in-house AI stack as Chinese handset makers push proprietary models into consumer software.

Huawei Cloud said in a Nov. 1 notice that its MaaS model platform will stop selling the openPangu-2.0-Flash and openPangu-2.0-Pro model services on Nov. 6, 2026, per IT Home, indicating a version transition. The report did not detail the new model's capabilities or benchmark results. Xiaoyi is Huawei's voice assistant, and the "Work" variant targets task automation on HarmonyOS. Huawei has steadily folded its Pangu model family into consumer and enterprise products; specifics on pricing or rollout were not disclosed.
Source: ithome.com · via Physix Frontier briefing desk · Editorial briefing card →

Tue, Nov 3, 2026

Bloomberg Tech segment weighs how AI giants will pay for compute buildout

A Bloomberg Tech video published Nov. 3, 2026 examines how large AI developers such as Anthropic and OpenAI plan to cover the cost of operating vast data-center capacity as they scale. The segment is framed as opinion and analysis rather than reported news.

The clip, titled "AI's Future Earnings Potential," poses the question of how AI companies will fund and monetize the computer warehouses underpinning their models. The available summary is truncated and does not disclose the participants' conclusions or any figures. Anthropic, based in San Francisco, develops the Claude models; OpenAI, also San Francisco-based, develops ChatGPT and the GPT series. Both have signed large cloud and chip supply agreements in recent years, though the segment's specific claims could not be verified from the provided excerpt. No financial details or company statements were included in the available material.
Source: bloomberg.com · via Physix Frontier briefing desk · Editorial briefing card →
OpenAI says ChatGPT search volume is closing in on Google

OpenAI said search query volume through ChatGPT is rising sharply, framing the chatbot as an emerging challenger to Google's search dominance, according to The Verge. The claim signals a shift in how some users find information online.

The announcement did not include specific query figures, and The Verge's summary of the post was truncated. OpenAI positioned the growth as evidence that "Ask Chat" is displacing "Google it" in some circles. Google remains the dominant search engine, but AI chatbots have begun absorbing some informational queries since ChatGPT launched in late 2022. OpenAI has been building out search features inside ChatGPT to compete directly. No independent data verifying the comparison was cited.
Source: theverge.com · via Physix Frontier briefing desk · Editorial briefing card →
OpenAI says ChatGPT hits 1.2 billion weekly active users

OpenAI consumer product lead Adam Fry said ChatGPT now has 1.2 billion weekly active users, up about 200 million in two months, according to an interview with Big Technology at a New York OpenAI event on Nov. 2.

The figure was reported by Chinese tech outlet IT Home on Nov. 4, citing Fry's remarks at the event. OpenAI has not published the number in a formal release. The 200 million gain over roughly two months would mark one of the fastest user expansions for a consumer AI product. ChatGPT remains the most widely used chatbot globally, ahead of rivals including Google's Gemini and Anthropic's Claude. IT Home's summary did not include further details from the interview, such as paid subscriber figures or regional breakdowns.
Source: ithome.com · via Physix Frontier briefing desk · Editorial briefing card →
AI and foldables fail to revive China phone market, BI says

Bloomberg Intelligence says AI features and foldable designs have not reignited Chinese smartphone demand, a signal for US chipmakers and component suppliers counting on a handset upgrade cycle in the world's largest phone market.

The research note, reported by Bloomberg on Nov. 3, did not disclose shipment figures or name specific vendors. Bloomberg Intelligence is the research arm of Bloomberg L.P. China's handset market is the largest single demand pool for Apple, Xiaomi and Oppo, and a sustained slump would weigh on orders for Qualcomm and MediaTek processors and memory suppliers. AI phone marketing has been a central pitch for Chinese vendors since 2024, with foldables positioned as a premium tier. BI's assessment suggests neither has moved replacement demand.
Source: bloomberg.com · via Physix Frontier briefing desk · Editorial briefing card →
Essay argues AI models will ship as torrents, citing DeepSeek release

A Hacker News-linked essay contends that large AI models will increasingly be distributed as torrents, pointing to DeepSeek's Oct. 17 release of its V4.1-Flash-1017 model as a 510 GB, 48-shard download. The claim is the author's argument, not a DeepSeek announcement.

The piece, published on pirateface.co, says the model was released as 48 safetensors shards plus a config and tokenizer, totaling 510 GB. It uses that packaging to argue torrent-style peer distribution will become standard for large weights. DeepSeek is the Hangzhou-based AI developer known for low-cost open-weight models. The essay does not say DeepSeek itself distributed the model via torrent; that framing is the author's. No independent confirmation of the release details was included in the source material.
Source: pirateface.co · via Physix Frontier briefing desk · Editorial briefing card →
Apple seeds iOS 27.3 beta with Watch AI listening tools

Apple has released a beta of iOS 27.3 that adds AI listening features for Apple Watch and a new health app, according to Bloomberg Tech, extending the company's on-device AI push into wearables.

The beta follows Apple's Worldwide Developers Conference, where the company previewed the Watch AI listening tools and the health app, Bloomberg reported on Nov. 3. The report also references China AI, though the input does not detail what that component involves. Apple has been working to bring AI features to its devices in China through local partners, a market where regulatory approval shapes which models can ship. Apple has not confirmed a public release date for the software.
Source: bloomberg.com · via Physix Frontier briefing desk · Editorial briefing card →
← NewerOlder →