China Desk

China News, Summarized 08 Oct 2026 34 stories archived day

Today's brief

China's toughest car safety regulator this week was a ByteDance app

Three brake pedals snapped on camera

Dongchedi broke the story and the stock. ByteDance's auto platform — more than 10m daily users, reportedly weighing a Hong Kong listing — published video on 8 October of three Zunjie V800 people-carriers losing their brake-pedal brackets during repeated 100km/h-to-zero stops. JAC, the Anhui carmaker that builds the brand with Huawei, hit its 10% limit down at 24.72 yuan ($3.48), shedding roughly 6.2bn yuan (~$873m) in a day, per TMTPost.

The forces involved were not heroic. On the third car, tested 4 October, the first stop logged a peak pedal force of 1,596 newtons over 38.44m; the bracket failed on the second stop at 1,612N, Jiemian reports. The viral "4,589N" screenshot was an artefact — a high-speed camera catching an LCD readout mid-refresh.

For scale, Qiu Jincao, deputy secretary-general of the advanced powertrain branch at the China Society of Automotive Engineers, posted the GM internal spec: 2,700N along pedal travel, 440N lateral, sustained, with no fracture and no deformation affecting return. An engineer told Jiemian the V800 bracket looks hand-built — reinforcement patched in here and there, a narrowed section near the pivot. Dongchedi says the fracture zone shows a suspected injection weld line.

Zunjie's statement is the interesting document. It claims development above the national braking standards and the industry pedal spec, reports no field failures, and calls the test "non-standard extreme conditions" — then promises to redesign the part and offer free upgrades to delivered owners, via IT Home. Conceding the fix while denying the defect. On a car priced 766,000 to 1,016,000 yuan ($108,000–143,000) that Yu Chengdong, Huawei's consumer chief, personally handed to a Hong Kong tycoon over the holiday.

WHY IT MATTERS: JAC booked 22.13bn yuan (~$3.1bn) of first-half revenue, up 14.31%, and still lost 749m yuan (~$105m). The stock had run up more than 40% in six sessions on the Huawei partnership. A third-party video repriced that in one session, and no regulator has said anything.

The day's other limit-down was a p-value. Allist Pharmaceuticals fell the full 20% to 89.26 yuan, losing over 10bn yuan (~$1.4bn) across the break, after its partner disclosed that furmonertinib missed first-line progression-free survival in exon-20-insertion lung cancer: hazard ratio 0.75, confidence interval crossing 1.02, p=0.0654, per TMTPost. The drug is 94.52% of Allist's revenue. The licence-out story was the other half of the valuation.

Go deeper on this section: ClaudeChatGPT

Sefcovic lands in Beijing with nothing to sign

Beijing published its terms before he arrived. The CCTV-linked account Yuyuan Tantian set out four things existing consultations could handle — joint monitoring of plug-in hybrid and chemical trade flows, price undertakings on the EV dispute, Chinese investment into Europe, steadier critical-mineral supply — and warned that staying ready to escalate mid-talks "cannot be called meeting each other halfway," translated by Sinocism.

The one thing Brussels wanted, it did not get. Trade commissioner Maroš Šefčovič opened two days of talks today having asked China to cap hybrid exports at roughly 15% of the EU market, against more than a third now. China refused before he landed. My read, following Monday's Macron–Merz letter: Beijing has decided Europe is bluffing. I am less sure than it is.

Meanwhile China put a name forward for the WHO. Beijing has nominated the health official Song Li for director general, which would make him the first person from the mainland to run the agency — as it absorbs the funding hole left by the US withdrawal, per SCMP. Vacancies at underfunded institutions are the cheapest influence on the market.

Go deeper on this section: ClaudeChatGPT

The West's open-weight comeback is scored in Chinese

Two Western labs shipped, and both chose Chinese yardsticks. Reflection's Beam — 501bn parameters, 23bn active, 23.8tn training tokens, four weeks of reinforcement learning on 10,500 GB300s — lands at 44.4 on an agentic coding benchmark against GLM-5.2's 44.0, but GLM-5.3 scores 61.0, Kimi K3 68.0 and DeepSeek V4.1 Flash 74.2, InfoQ reports. Mistral's Large 4 beats DeepSeek V4 Pro and loses to Kimi K3.

The pitch has shifted from best to cheapest-per-unit-of-smart. Reflection concedes Kimi K3 leads outright and sells inference efficiency instead — which is DeepSeek's own 2025 argument, made with American capital and Nvidia's newest silicon. Chinese models now account for about 41% of Hugging Face downloads, and open weights handled 56% of Vercel's gateway tokens in August against under 10% last December.

Huawei's Eric Xu offered the mirror image. The rotating chairman argues American labs feel AI risk more acutely because they have far more compute and may have seen capabilities Chinese developers have not; China "still needs to accelerate," and may come to share those concerns later, per Pekingnology. Risk perception as a function of FLOPs.

Manus came home with half a billion dollars. Butterfly Effect, which moved to Singapore in May last year and cut its roughly 120-person China team in July, has raised over $500m led by Boyu and IDG, and is hiring 17 roles in Beijing — Agent harness, Agent evaluation, LLM algorithms, per QbitAI. Founder Xiao Hong got the company back from a blocked Meta sale in September. Whether engineers he laid off fifteen months ago take the call is the actual story.

Go deeper on this section: ClaudeChatGPT

Two service-labour stories landed on the same day

An airline association sided with its own cabin crew. After video spread of an Eastern Airlines flight attendant kneeling to apologise, the China Air Transport Association said a meal cart brushed a passenger, the crew apologised repeatedly, and the passenger kept up verbal attacks and threats — the airline has preserved evidence and filed a police report, per 36Kr's roundup. The association called the conduct a suspected offence, not a service failure.

Then a writer described two weeks inside Sam's Club. The first-person account in TMTPost: 27.7 yuan (~$3.90) an hour in Beijing, four-hour shifts, 40 to 50 members an hour through a single exit on a weekday afternoon, bathroom breaks by group-chat permission, meal breaks clocked to exactly 45 minutes, two weeks' notice for any leave.

The retention mechanism is the tell. Health-certificate costs are only reimbursed after three months; quitting early means a permanent blacklist covering Walmart too, the essay says. Against 10.7m paying members and a wall motto about happy employees that the writer says nobody quotes internally any more. Both stories are about the same thing: who absorbs the cost of Chinese service standards.

Go deeper on this section: ClaudeChatGPT

Everyone led with a different half of the Korean bank hack

The breadcrumb was a résumé. Seven South Korean financial groups were breached. Solidot's account foregrounds the self-own: the attacker used Claude Code to write himself a security-researcher CV showcasing his intrusions, typing in his initials, Telegram handle, education and city — and stresses that CrowdStrike's identification of a 26-year-old in Guangdong is inference, not established.

Western coverage led with the nationality. The Register's headline on the same report puts "suspected Chinese speaker" and the Chinese-built ARTEX pentest agent up front.

Both buried the stack. The ARTEX instance ran DeepSeek V4.1-Flash as its primary model, with Zhipu's GLM-5.3 and Grok 4.6 in other sessions, per the report. Claude Code was the console. Chinese open weights were the engine — which is precisely the diffusion risk Anthropic's own misuse report described, demonstrated against it.

Go deeper on this section: ClaudeChatGPT

Threads we are pulling

  • The game studios were DeepSeek's quiet LPs. Following Monday's piece on the round now nearing 100bn yuan (~$14bn): fifteen listed Chinese game companies put over 87.2bn yuan (~$12.3bn) into 72 deals since early 2025, under 1% of it in entertainment. NetEase wrote about 3bn yuan and Tencent about 10bn into DeepSeek's June round, per TMTPost. 37 Interactive's 0.27% of Zhipu booked 1.016bn yuan (~$143m) of first-half investment income — 45.22% of total profit.
  • The Claude access thread keeps cutting both ways. Yesterday I noted payment failures; today V2EX posters report the $20 Pro tier banned outright and renewals blocked after 8 November. Sentiment, not data — but the same week Claude Code became a forensic exhibit.
  • The holiday conversion gap has a physical cap. Following the tourism numbers: Shandong's "follow the commander and storm the county town" reenactment seats 1,200 people at a site that took 75,000 visitors a day last National Day, TMTPost reports. Datong's rebuilt Ming prince's mansion did 640,000 visitors and 33.9m yuan (~$4.8m) last year throwing foam rocks. Scarcity is the product.
  • The slop thread got a second failure mode. Leiphone ran four more instalments of that anonymous 188-volume AI-governance codex plus a full table of contents. QbitAI, meanwhile, published a page reading only "hacked by trenggalek6etar". One masthead is filling with machine output; the other cannot hold its CMS.

Go deeper on this section: ClaudeChatGPT

The river · 34

Porsche plans deep cuts as China sales keep falling

# QbitAI

Porsche plans to cut up to 30% of its workforce, shrink management by 40% and revive combustion-engine investment after profit margin fell from 18% to 1.1%.

More
  • The source says China deliveries fell from 95,671 in 2021 to 41,938 in 2025, then dropped another 32% year over year in the first half of 2026.
  • Porsche is extending combustion and hybrid product lifecycles while simplifying models, a reversal of its earlier plan for more than 80% electric sales by 2030.
  • The strategic problem is not only weak EV demand: Chinese premium brands now combine luxury positioning with software, intelligent cabins and electric performance.
  • The company plans to cut about 9,000 jobs through attrition, retirement and voluntary departures while investing 2.1 billion euros, roughly $2.3 billion, in German plants and R&D.
  • The case shows how China’s premium EV competition is forcing foreign brands to trade technology ambition for financial resilience.
Open at QbitAI Ask Claude ChatGPT

Manus rebuilds a Beijing team for China-facing agents

# QbitAI

After moving its center abroad and surviving a blocked acquisition, Manus is hiring 17 Beijing roles across agent engineering, evaluation, security, product and growth.

More
  • The hiring plan suggests a full local business unit rather than a small research outpost, with roles spanning the execution harness, evaluation, infrastructure and commercialization.
  • The company says it is developing a domestic product and working with Chinese model providers and ecosystem partners.
  • This reversal shows how ownership, regulation and market access can pull a Chinese AI company back after an internationalization attempt.
  • Recruiting after a prior China-team retrenchment may be difficult: candidates will assess whether the domestic product is a durable strategy or another temporary reorientation.
  • The round’s reported $500 million-plus financing gives the reset resources, but no launch date or domestic product metrics are available.
Open at QbitAI Ask Claude ChatGPT

Huawei says China may confront AI risk as its capabilities catch up

# Pekingnology

Huawei chairman Eric Xu says China must accelerate AI now but may share US developers' safety concerns once its capability frontier reaches the same point.

More
  • Xu's framing is unusually explicit: perceived risk is partly a function of how far a country's models and compute have advanced, not simply a matter of abstract principle.
  • He couples acceleration with risk management, arguing that China should develop AI while ensuring it is used for beneficial rather than harmful purposes.
  • The same discussion reportedly covered Huawei's interconnect and computing architecture, reflecting the company's attempt to build a full stack under advanced-chip restrictions.
  • The political implication is that Chinese AI policy may remain expansionary until domestic systems encounter risks that US frontier labs already describe, rather than adopting a matching slowdown now.
  • The statement also reveals how Chinese technology leaders interpret US safety rhetoric: partly as a genuine capability response and partly as a condition China has not yet fully experienced.
  • For executives, the practical signal is continued Chinese investment in compute efficiency, networking and domestic alternatives despite export controls.
  • The report is based on a Chinese media transcript and offers Huawei's position, not an official national policy statement.
  • Whether this posture changes will depend on domestic capability milestones, incident experience and how Beijing balances safety with strategic competition.
Open at Pekingnology Ask Claude ChatGPT

Alibaba signals a larger, cheaper next generation of Qwen

# Leiphone

At its cloud conference, Alibaba said future Qwen models may scale toward 5T to 10T parameters while using sparse attention, memory and selective activation to control inference cost.

More
  • The reported Qwen3.8 model has 2.4T total parameters but activates about 6B per inference in the Flash variant, illustrating scale without dense compute at every token.
  • Alibaba says the architecture cut training cost to one-ninth, input pricing to one-quarter and long-context prefill time by 8.6 times versus its predecessor.
  • The company is also using models to design chips and optimize infrastructure; one reported NoC experiment ran for more than 60 hours and cut estimated area 42% and power 59.5%.
  • This is a concrete Chinese version of recursive self-improvement: the model assists engineering inside a human-defined workflow, rather than independently choosing the company's goals.
  • The strategic bet is that China can offset hardware constraints through architecture, compiler and systems co-design while continuing to scale foundation models.
  • The figures are Alibaba disclosures and need independent reproduction, but they show the competitive frontier moving from parameter counts toward active compute and engineering loops.
  • The next Qwen releases will reveal whether sparse architectures preserve quality at very large scale and whether Alibaba can turn internal optimization into a durable cloud advantage.
Open at Leiphone Ask Claude ChatGPT

Western open models accelerate as Chinese models set the benchmark

# InfoQ China

Reflection and Mistral launched open-weight models against Chinese systems, while Chinese models reportedly account for 41% of Hugging Face downloads and 56% of Vercel gateway tokens.

More
  • Reflection’s Beam is a 501-billion-parameter MoE model with 23 billion active parameters, trained on 23.8 trillion tokens and more than 10,000 GB300 GPUs.
  • Beam reportedly matches an older Chinese model on some coding tests but trails newer systems, shifting its pitch from absolute leadership to intelligence per unit of inference compute.
  • The open-weight market is moving from research novelty to production infrastructure, making download share and token share strategically meaningful even when benchmark rankings move quickly.
  • The Chinese framing treats DeepSeek, Qwen, Kimi and GLM as the reference class against which Western open models must now compete.
  • The underlying contest is not simply openness; it is who can deliver a durable capability-to-cost curve with usable tooling and supply.
Open at InfoQ China Ask Claude ChatGPT

China's autonomous freight fleets move toward mixed convoys

# Leiphone

Kardrive says it has reached route-level profitability with autonomous heavy-truck convoys and is testing cabless freight robots across six provinces and major energy corridors.

More
  • The company reports more than 100 vehicles, 30 customers and operations spanning Inner Mongolia, Xinjiang, Gansu, Shaanxi and Shanxi, where repetitive mine-to-plant routes make autonomy easier to constrain.
  • Its cost model identifies tolls and energy as roughly 30% each, drivers 20% and maintenance and depreciation 10% each; convoying attacks the driver component first.
  • Reported operations show a 10% to 18% gross-margin improvement over conventional transport, while cabless vehicles could raise vehicle economics from about 20% to above 30%.
  • The convoy is not simple platoon following: the rear truck retains independent perception and control for cut-ins, traffic lights and network dead zones, while sharing lead-vehicle information as an extra signal.
  • The proposed network mixes human-led trucks, cab-equipped autonomous trucks and cabless robots, shifting the problem from single-vehicle autonomy to dispatch, maintenance, cargo demand and route operations.
  • The company says route deployment time has nearly halved from eight to twelve months, but the excerpt gives no audited fleet-level uptime or safety record.
  • The strategic test is whether a profitable constrained corridor can become a repeatable logistics network without recreating the labor and operational overhead it seeks to remove.
Open at Leiphone Ask Claude ChatGPT

After the Xi-Trump summit, Beijing seeks status while Washington seeks deals

# SCMP

Chinese coverage reads the summit as a question of equal standing and recognition, while the US framing emphasizes reciprocity and transactional outcomes.

More
  • The analysis treats ceremony, seating and public reception as strategic signals about rank rather than diplomatic decoration.
  • Beijing's preferred outcome appears to be recognition as an equal pole in negotiations, not merely a list of concessions exchanged under pressure.
  • The US framing described here is more deal-oriented, leaving unresolved whether reciprocal bargaining also implies equal status.
  • That ambiguity matters to other countries deciding whether the relationship is a condominium, a rivalry managed by transactions or a hierarchy still controlled by Washington.
  • A successful summit can lower immediate escalation risk without resolving the underlying conflict over technology, security and regional order.
  • The story is interpretive and does not establish a signed settlement beyond the source's reference to an agreement built around reciprocity.
Open at SCMP Ask Claude ChatGPT

Manus raises more than $500 million at a reported $4 billion valuation

# Solidot

Manus completed one of China’s largest AI application financings after its failed Meta sale, with Chinese and dollar-fund investors backing its independent restart.

More
  • The financing was led by Boyu Capital and IDG, with Tencent, Sequoia China and ZhenFund returning as investors.
  • The reported $4 billion post-money valuation makes Manus one of China’s most valuable agent startups, but the figure is not supported by a public filing.
  • The company’s regulatory reversal shows that ownership of strategic AI firms can be constrained by Chinese authorities even after a foreign acquisition is announced.
  • The restart gives Manus capital and autonomy, but it must now demonstrate recurring usage and defensible infrastructure rather than rely on the scarcity value of the deal itself.
Open at Solidot Ask Claude ChatGPT

Brake-bracket test sends JAC shares to a 10% limit fall

# TMTPost

A third-party emergency-braking video showed three Fangjie V800 vehicles with broken pedal brackets, wiping roughly 6.2 billion yuan, about $870 million, from JAC’s market value.

More
  • Fangjie says the test was an extreme nonstandard condition and that delivered vehicles had no reported failures, but it will redesign the bracket and offer a free upgrade.
  • The response avoids admitting a regulatory defect while acknowledging that additional safety margin is commercially necessary for a premium vehicle.
  • The market move also reversed part of a prior rally tied to Huawei cooperation and high-end vehicle expectations, so the loss was not caused by one test alone.
  • For Chinese automakers, the episode shows how quickly social-video evidence can challenge engineering validation and brand trust.
  • The unresolved questions are the exact structural change, validation boundary and retrofit execution; those will determine whether the controversy fades or becomes a quality-system issue.
Open at TMTPost Ask Claude ChatGPT

Huawei says more phones will use its LogicFolding chips

# IT Home

Huawei executives say vertical logic stacking will spread across future phones as China adapts chip architecture to restricted access to advanced nodes.

More
  • The reported LogicFolding approach vertically stacks logic layers and uses dense interconnects to shorten critical paths and reduce latency; the source gives no independent process or yield data.
  • Huawei says its latest mobile chip integrates internally designed CPU, GPU, NPU, modem, ISP and security blocks, a strategy aimed at system-level control rather than node parity alone.
  • The proposed trajectory toward 400-plus million transistors per square millimeter and 5 GHz by 2031 is a roadmap claim, not a demonstrated milestone.
  • The strategic signal is clear: Huawei is framing architectural integration and packaging as substitutes for unrestricted access to leading-edge manufacturing.
Open at IT Home Ask Claude ChatGPT

China's AI infrastructure is shifting from peak compute to token economics

# TMTPost

A Chinese industry analysis says inference demand is moving the infrastructure race from training peak performance toward reliable, low-cost token production and tightly coupled supernodes.

More
  • The article cites Chinese token-use estimates rising from 100 billion daily tokens in early 2024 to more than 140 trillion daily tokens in March 2026, though the measurement methods are not explained.
  • It argues that inference changes the purchasing metric from accelerator peak speed to sustained tokens per second, power and depreciation per token, utilization and failure recovery.
  • The proposed hardware response is to expand high-bandwidth interconnect domains from eight-GPU servers to 64 or more accelerators, making a logical supernode rather than a loose rack of machines.
  • That architecture targets communication bottlenecks in tensor and expert parallelism, where moving activations and weights can leave expensive compute idle.
  • The piece cites aggressive forecasts from banks and IDC, but its strategic conclusion is more robust than any one forecast: agent workloads make inference a persistent operating expense.
  • For Chinese vendors, this favors integrated servers, networking, cooling and scheduling systems over isolated accelerator specifications.
Open at TMTPost Ask Claude ChatGPT

Chinese automakers push deeper into battery design and manufacturing

# TMTPost

Xiaomi, Li Auto, Leapmotor and XPeng are expanding battery self-development to regain control over cost, vehicle architecture and supply risk rather than simply abandoning suppliers.

More
  • Battery packs now shape vehicle stiffness, cabin space, thermal behavior, fast charging and software-controlled range, making the old buyer-supplier boundary strategically untenable.
  • The reported programs differ: Xiaomi co-defines cells and packs with suppliers, XPeng emphasizes self-developed packs and diversified cell sources, while Leapmotor claims end-to-end cell and pack production.
  • Li Auto is taking a deeper equity route, reportedly investing another 26.5 billion yuan (roughly $3.7B) in Sunwoda Power and forming a 50-50 battery venture.
  • The common objective is bargaining power and product definition, often through supplier investment and dedicated lines rather than fully internalizing every electrochemical process.
  • This is a capability race in systems integration, manufacturing control and data access; the label self-developed battery should not be read as complete cell independence.
  • Success will depend on yield, safety, warranty performance and the ability to scale new chemistry without disrupting vehicle deliveries.
Open at TMTPost Ask Claude ChatGPT

China blocks an offshore restructuring of Manus, then lets it go independent

# IT Home

Manus parent Butterfly Effect raised more than $500 million (roughly $70M), after Chinese regulators blocked an earlier Meta transaction and required the company to restore its pre-deal ownership structure.

More
  • The reported round is led by Boyu Capital and IDG Capital, with Tencent, Sequoia China and ZhenFund continuing as investors.
  • The decisive event is regulatory rather than financial: China's foreign-investment security-review office reportedly prohibited the proposed acquisition and rejected a restructuring that regulators viewed as improper offshore relocation.
  • The report says the company moved its headquarters to Singapore, cut its mainland team and stopped domestic service, triggering debate over whether Chinese engineering and infrastructure could be separated from ownership and control.
  • After the blocked transaction, the founders and original investors reportedly repurchased all Meta shares at a $2 billion valuation, with Tencent taking the former Benchmark stake.
  • Manus later announced a split from Meta and a return to independent operation, illustrating Beijing's willingness to unwind a cross-border deal without necessarily shutting down the underlying product.
  • The broader signal is that an offshore corporate shell does not automatically remove a Chinese AI asset from national-security and capital-control scrutiny.
Open at IT Home Ask Claude ChatGPT

China is building compute capacity at a pace that could reshape AI economics

# Techmeme

SemiAnalysis estimates China has 24 GW of operational compute and another 50 GW planned or under construction, versus 56 GW operating in the US.

More
  • The figures are an aggregator's summary of Financial Times coverage, not a Chinese government capacity census, so definitions of compute and inclusion criteria matter.
  • Even with that caveat, the planned pipeline suggests Beijing is treating AI infrastructure as an industrial buildout rather than a scarce cluster of research facilities.
  • The geographic reference to Inner Mongolia points to China's use of low-cost power and large industrial sites for data-center expansion.
  • Capacity alone does not reveal useful throughput: accelerator mix, networking, utilization, model access and electricity availability determine how much intelligence the fleet can actually deliver.
  • If the estimate is directionally right, export controls may constrain individual chips without preventing China from pursuing scale through domestic systems and massive deployment.
Open at Techmeme Ask Claude ChatGPT

PaperBenchX finds AI systems rarely reproduce full scientific papers

# AI QbitAI

A new 93-task benchmark reports that the best tested model fully reproduced only 13.98% of scientific papers, despite stronger scores for modeling and execution.

More
  • PaperBenchX hides the scoring criteria, runs submitted workflows in an isolated offline environment and scores regenerated evidence rather than plausible answer numbers.
  • The benchmark spans 12 fields and 3,168 expert-checked criteria, exposing a gap between producing outputs and proving that a scientific workflow is valid.
  • Average modeling and execution scores were about 62%, while validation was only 42.8%, locating the bottleneck in verification rather than raw tool use.
  • This is a technically important distinction for AI-for-science systems: reproducibility, provenance and failure diagnosis matter more than benchmark answers that can be guessed from papers.
  • The benchmark is new and partially closed, so its generality and resistance to contamination still need independent scrutiny.
Open at QbitAI Ask Claude ChatGPT

Manus secures a record domestic agent round

# AI Zhidx

Chinese coverage says Manus raised more than $500M, or about 3.35B yuan (roughly $470M), at a reported $4B valuation after resuming independent operations.

More
  • The round is described as the largest single financing for a Chinese agent startup, with Boyu and IDG leading and Tencent, Sequoia China and ZhenFund following.
  • Manus 2.0 uses its Cascade framework, which reportedly reduced token use 23.2%, task time 28.2% and operating cost 32% in one test.
  • The financing follows the blocked Meta acquisition and the company's return to independent operation, making ownership and geopolitical constraints part of the business story.
  • The capital arrives as global model providers and open-source projects push agents down into platforms, so Manus must defend a product and distribution layer rather than merely demonstrate autonomy.
  • The valuation is based on investor financing, not public revenue or retention, and the report gives no evidence that Cascade's test improvements generalize across customers.
  • The event shows Chinese capital treating execution-oriented agents as a strategic category even after regulators redirected the company's cross-border path.
Open at Zhidx Ask Claude ChatGPT

Claude Haiku 5.5 cuts price while keeping a coding ceiling

# AI Zhidx

Anthropic’s Haiku 5.5 reportedly cuts average cost about 75% and improves computer use, but remains far behind Sonnet on complex terminal coding.

Also: Leiphone, QbitAI

More
  • The model supports adjustable effort, with the source citing OSWorld performance from 42% at low effort to 72.4% at maximum effort.
  • The savings depend on context length and tokenizer behavior; code, tables and non-English text may consume more tokens than headline pricing suggests.
  • Terminal-Bench performance of 39.2% versus Sonnet’s 70.6% makes the deployment boundary clear: use Haiku for bounded subtasks and Sonnet or Opus for long-horizon coding.
  • API changes include adaptive thinking, altered computer-use interfaces and removal of some sampling controls, creating migration work for existing applications.
  • Chinese coverage frames the release as a price war with DeepSeek and other open models, but real task cost remains more informative than list price.
Open at Zhidx Ask Claude ChatGPT

Geely claims 2.25 MW charging and AI battery management

# Analysis QbitAI

Geely unveiled a charging system claiming 2.25 MW peak power, sub-65 C battery temperature and a 20% cycle-life improvement through AI control and liquid cooling.

More
  • The architecture combines battery, charger, vehicle, grid and storage data, with a model predicting internal temperature changes up to 30 seconds ahead.
  • A five-part liquid-cooling path covers storage, power conversion, cable, battery pack and vehicle inlet; Geely says immersion cooling improves converter heat transfer by more than 30%.
  • The company reports a 10% to 97% charge in 8 minutes 40 seconds on one vehicle, but independent fleet durability and grid-capacity evidence is absent.
  • The strategic move is to make charging infrastructure and battery health part of the vehicle platform, not a separate utility experience.
  • The headline power figures are impressive but commercially meaningful only if stations, connectors and local distribution networks can support them.
Open at QbitAI Ask Claude ChatGPT

Shanghai HiSilicon says edge AI will win through developer ecosystems

# Analysis Zhidx

Shanghai HiSilicon argues that China's next AI wave will be built into everyday devices, and that chip vendors must provide tools, models and hardware support rather than silicon alone.

More
  • The company frames edge AI around latency, sensor access, privacy and offline operation, with cloud and device models forming a reinforcing loop.
  • Its central commercial insight is that small firms can buy compute but still lack the embedded engineering needed to balance memory, power, cost, quantization and device-specific kernels.
  • That makes SDKs, BSPs, demos, debugging tools and developer communities strategic infrastructure, not marketing extras.
  • The article's examples of smart toothbrushes and hair dryers are illustrative rather than deployed products, but they show the breadth of the proposed light-intelligence category.
  • The strategy reflects China's fragmented IoT market, where a chip supplier needs to reduce integration work across many small manufacturers.
  • The real test will be whether HiSpark converts developers into repeat production deployments rather than isolated demos.
Open at Zhidx Ask Claude ChatGPT

A Chinese startup pitches 24/7 robot-staffed convenience stores

# AI QbitAI

Zhengxing Innovation unveiled humanoid and wheeled-arm robots plus a management platform for stocking, inspection and transport in open convenience-store environments.

More
  • The system is designed for zero store remodeling, using a narrow omnidirectional base for aisles and a humanoid for customer-facing work.
  • Its M1 platform assigns tasks, monitors state and supports human takeover across multiple robot types, making fleet operations as important as the bodies.
  • The company claims 99% autonomous task completion, 98.6% success on LIBERO and a 24-times inference-speed advantage for its lightweight world-action model.
  • The planned 2027 business model includes direct sales and robotics-as-a-service, with hardware, software, maintenance and upgrades bundled together.
  • Retail is a harder deployment environment than a warehouse because customers, shelves, SKUs and layouts change continuously.
  • The figures are company claims and the announced commercial partnership is still a validation program, so labor savings and incident rates remain unknown.
Open at QbitAI Ask Claude ChatGPT

A Chinese robot learns precision CNC handling with 600 trajectories

# AI Leiphone

Simplexity Robotics reports using 600 real robot trajectories to handle CNC loading with only a 0.5-millimeter clearance, targeting the gap between lab benchmarks and factory tolerances.

More
  • The system must handle reflective metal, weak visual texture, similar parts and contact interfaces that are not always visible while tending two CNC machines.
  • Its architecture combines a world-action model for candidate motion, a fixed-size recurrent memory for long context and a separate evaluator that scores candidate action chunks.
  • Training reconstructs future video, 3D state and actions, but inference predicts actions directly without rendering the future, reducing the runtime path.
  • The report cites 15 successful trials for the fused method versus zero for one comparison and eight for a supervised baseline, though the sample is small.
  • The engineering lesson is that factory deployment needs memory, action selection and contact-aware perception in addition to a multimodal backbone.
  • A 0.5-millimeter gap makes tolerance, cycle time and yield first-class metrics; a polished demo that ignores these constraints would not translate to manufacturing.
  • The claimed robot-making-robots feedback loop is strategically attractive, but sustained multi-shift operation and quality control remain unproven.
Open at Leiphone Ask Claude ChatGPT

Huawei and Nvidia treat inference KV cache as a storage tier

# AI Leiphone

Huawei and Nvidia are developing dedicated storage for reusable inference KV cache, but the SSD endurance and capacity standard has not stabilized.

More
  • Externalizing KV cache trades storage cost for recomputation: raising hit rate from 90% to 95% halves the portion that must be recomputed, but may require disproportionate capacity.
  • DRAM can cost dozens of times more than enterprise SSDs, while some designs may need five to ten times as much SSD capacity for equivalent cache capacity.
  • Workload lifetime is the unresolved variable: minute-scale caches create write pressure, while week-scale caches emphasize capacity and retrieval efficiency.
  • The source says current customers still use ordinary enterprise SSDs and that a common product standard may take six to twelve months, followed by roughly another year of development.
  • This is an important infrastructure signal because inference memory is becoming a systems-design problem rather than a server-component detail.
Open at Leiphone Ask Claude ChatGPT

Google uses Gemini and fuzzing to replace a C library with Rust

# AI InfoQ China

Google translated a 3,000-line C library into ABI-compatible Rust, using differential fuzzing and 30 million real files to validate the replacement.

More
  • The migration preserved exported symbols and data structures, while human engineers still had to review raw-pointer ownership and lifetime rules at the FFI boundary.
  • Two hundred million fuzzing iterations over six days found behavioral issues and an old out-of-bounds write, illustrating that generated code was only one part of the safety case.
  • The practical lesson is architectural: memory safety moved into the type system, allowing Google to remove a process sandbox without increasing latency.
  • The approach is most credible for small, self-contained libraries; larger dependencies still face semantic drift, upstream fork maintenance and unsafe-interface review.
  • Chinese coverage presents the project as a model for AI-assisted systems migration, while the cited developer discussion remains skeptical that one-shot translation scales.
Open at InfoQ China Ask Claude ChatGPT

Google puts Gemini Agent at the center of enterprise work

# AI Zhidx

Chinese coverage frames Gemini Agent as a general-purpose digital coworker that can route among Gemini and Claude, maintain four kinds of memory and continue work after the user leaves.

More
  • The product combines goal-based planning, enterprise data access, code execution, document creation and multi-agent delegation behind one work interface.
  • Its four memory types separate current context, business knowledge, procedures and past episodes, a more structured design than simply appending old chats to a prompt.
  • The coworker concept gives an agent an email address, calendar, storage and directory identity, making permissions and auditability more consequential than in a personal chatbot.
  • Google says routing can cut total cost to one-third of an all-frontier-model baseline and improve speed by 40%, but those figures are company claims.
  • The Chinese framing places Gemini Agent alongside domestic workplace agents such as Doubao Work and Qianwen, revealing that the competitive unit is becoming the enterprise operating model rather than the base model.
  • The unresolved issue is governance: persistent identity, long-running execution and cross-application permissions create a much larger failure surface than answer generation.
Open at Zhidx Ask Claude ChatGPT

A Tsinghua-founded startup bets on action-native robot learning

# AI Leiphone

BASAL Intelligence, founded by a first-generation Tsinghua AIR doctoral graduate, is pursuing action-native in-context learning so robots can adapt to new tasks from a few demonstrations instead of retraining.

More
  • The proposed model treats actions, feedback, failures and demonstrations as the primary representation, with long context used to update behavior after deployment.
  • Its X-VLA model targets transfer across robot bodies, while ODEWorld models continuous physical time to support long-horizon in-context learning.
  • The team says X-VLA used fewer than 1,000 hours of pretraining data and won an IROS 2025 robot challenge, but these claims do not yet demonstrate production robustness.
  • The commercial thesis is powerful: if a robot can learn in minutes on site, deployment cost could grow sublinearly with the number of environments and tasks.
  • The non-language-centered approach will face integration friction because customers and current VLA stacks are organized around verbal commands and post-training.
  • The key validation is not another benchmark but repeatable adaptation across new bodies, objects and workflows without hidden fine-tuning.
  • Funding details are not disclosed, so the story is primarily a technical and talent signal rather than a financing event.
Open at Leiphone Ask Claude ChatGPT

OpenAI productizes multi-agent systems as researchers question scale

# AI InfoQ China

OpenAI launched collaborative Dots and an Agents API, while researcher Noam Brown says current evidence does not show that more agents create proportional gains.

More
  • Brown describes multi-agent systems as parallel test-time compute: four or sixteen agents can improve speed, but coordination overhead makes scaling sublinear and task-dependent.
  • The cited architecture lets agents message one another directly rather than relying only on a fixed coordinator-and-worker hierarchy, increasing flexibility at the cost of control complexity.
  • The strategic point is that stronger base models and longer inference may matter more than agent count; the source says less than 10% of a benchmark breakthrough should be credited to multi-agent organization.
  • As products move from demos to persistent workspaces, communication, context sharing, safety and goal alignment become the bottlenecks.
  • This is a useful counterweight to the industry’s agent-count marketing, but the claims remain interview analysis rather than a full controlled study.
Open at InfoQ China Ask Claude ChatGPT

US proposes a $70,000 annual fee for F1-OPT work authorization

# Policy Solidot

A proposed US rule would raise the F1-OPT fee to $70,000 and STEM extensions to $30,000, effectively ending post-study work for many international graduates.

More
  • The proposal targets the work authorization route commonly used by international students after graduation, with public comments due November 9.
  • Chinese students and technology employers would be heavily exposed if adopted, making this a direct talent-mobility shock rather than a routine visa change.
  • The report comes from Chinese media and presents the proposal as job protection, but the supplied text does not include the rule’s legal rationale or implementation risks.
Open at Solidot Ask Claude ChatGPT

Naive AI raises $400M in seven months, then stays invisible

# AI Leiphone

A secretive Tsinghua-linked startup reportedly raised $400M across three rounds at a $1.42B valuation, showing how Chinese AI capital can fund teams before public product disclosure.

More
  • Naive AI is founded by Tsinghua associate professor Dai Jifeng and reportedly brings together eight researchers with backgrounds at Microsoft Research Asia and SenseTime.
  • The team includes people associated with deformable convolution, video recognition and AI infrastructure, giving the company unusually deep technical lineage for a young startup.
  • The seven-month financing pace and sparse public information suggest investor conviction is being built around talent, prior research and private diligence rather than visible product traction.
  • The report says the team has not accepted interviews and offers no product description, customers or revenue, making the valuation impossible to evaluate operationally.
  • This is a useful signal about China's current capital market: top research teams can command enormous funding before they expose a model or application, but that also concentrates execution risk.
  • The important future test is whether the group produces a differentiated system rather than simply recreating the research-lab prestige that attracted the money.
Open at Leiphone Ask Claude ChatGPT

A Chinese state-property group absorbs a troubled Minmetals developer

# Policy 36Kr

MCC Property is acquiring 100% of Minmetals Real Estate within the China Minmetals system, consolidating assets while taking on a target with negative net assets and 125% leverage.

More
  • The transaction is an internal state-owned-enterprise restructuring rather than a market takeover, designed to centralize property management and professionalize operations.
  • Minmetals Real Estate reportedly had 19.87B yuan (about $2.8B) of assets, 24.86B yuan (roughly $3.5B) of liabilities and negative net assets of 4.99B yuan (about $703M) at midyear.
  • The transfer follows a 2026 delisting and an 8.8B yuan (about $1.24B) capital injection used to repay overseas borrowing, showing how the group is cleaning up the vehicle before or during consolidation.
  • A separate urban-renewal unit has positive net assets and 31 projects across 20-plus cities, so the deal mixes distressed liabilities with potentially useful operating assets.
  • The transaction illustrates how central state-owned groups use internal transfers to contain property-sector stress without a conventional bankruptcy or open-market sale.
  • The forecast combined balance sheet still shows a 74.94% debt ratio and a projected 624M yuan (about $88M) net loss, so ownership change is not the same as economic repair.
  • The important question is whether the parent can extract assets and talent while ring-fencing legacy debt, rather than merely moving the liabilities on paper.
Open at 36Kr Ask Claude ChatGPT

IROS awards a long-term memory system for robots

# AI Leiphone

IROS 2026’s best paper introduced volatility-aware spatial-temporal memory that separates current state, changes and long-term patterns for robots living in dynamic environments.

More
  • The LT-Mem design uses Live Memory, Delta Memory and Meta Memory, with object volatility determining how history is retained and queried.
  • The award signals a shift from adding model scale toward handling months or years of environmental change, an essential problem for service robots.
  • Other highlighted work combined physics-informed residual learning with tight drone formation control and demonstrated humanoids playing tennis or carrying trays.
  • The Chinese analysis uses the awards to argue that traditional geometry, dynamics and control remain central even as foundation models enter robotics.
Open at Leiphone Ask Claude ChatGPT

Qi Anxin leads China's cybersecurity market for a seventh year

# Signal Leiphone

A CCID report puts China's cybersecurity market at 93.08B yuan (roughly $13.1B) in 2025 and ranks Qi Anxin first overall, with 43.9B yuan (about $6.2B) in revenue.

More
  • Qi Anxin reportedly held 17.5% of endpoint security and 18.3% of security-management platforms, while also ranking first in security services.
  • The market is shifting from new construction toward optimization, platform integration and managed services, making a broad product stack more valuable than isolated appliances.
  • The company's AISOC combines a security model with a data-correlation engine; it claims sub-nine-second alert triage and over 93% accuracy in one luxury-automaker deployment.
  • Qi Anxin also reports more than 95% alert-noise reduction and AI-assisted automation across much of its penetration-testing workflow, but these are vendor metrics.
  • The report is useful as a market-structure signal: Chinese buyers are consolidating endpoint, network, data and service operations into integrated domestic platforms.
  • The missing information is methodology—market definitions, competitor shares, contract concentration and independent validation of the AI metrics.
Open at Leiphone Ask Claude ChatGPT

Chinese researchers claim a cheaper route to sustainable aviation fuel

# Signal SCMP

A Chinese team says a cobalt-manganese catalyst converted biomass-derived synthesis gas into jet fuel in a 1,000-ton pilot, with modeled gross margins above 50%.

More
  • The process combines syngas from woody biomass with a catalyst designed for efficient jet-fuel synthesis, linking a chemical route to lower-carbon aviation fuel.
  • A thousand-ton pilot is more meaningful than a lab result, but it is still far below commercial plant scale and does not establish lifecycle emissions.
  • The claimed economics depend on biomass cost, syngas preparation, catalyst life, product separation and carbon accounting, none of which are detailed in the supplied report.
  • If reproducible, the advance could be strategically important for China because it connects domestic biomass and chemical-engineering capability to an aviation fuel bottleneck.
  • The item is a hard technical and economic signal, but independent process data and a full techno-economic model are needed before treating the 50% margin as investable.
Open at SCMP Ask Claude ChatGPT

IROS papers show robotics resisting foundation-model monoculture

# AI Leiphone

A review of 1,933 IROS papers finds learning, planning, perception, control and manipulation all expanding alongside foundation models rather than being replaced by them.

More
  • The review counts about 809 robot-learning papers, 564 on navigation and planning, 556 on perception and vision, 546 on control and dynamics, and 520 on manipulation.
  • Cross-paper analysis shows language models being embedded inside geometric planning, 3D representations, tactile sensing and fast policy execution rather than directly replacing those modules.
  • The systems implication is architectural: slow, powerful models can teach or guide smaller real-time controllers, while geometry and contact remain hard constraints.
  • The review is a media analysis of conference papers, not a peer-reviewed meta-study, but it captures a meaningful shift from end-to-end enthusiasm toward layered systems.
Open at Leiphone Ask Claude ChatGPT

China nominates Song Li to lead the World Health Organization

# Policy SCMP

China nominated health official Song Li for WHO director general, presenting the bid as a chance for mainland China to lead the agency for the first time.

More
  • The nomination comes as the WHO faces a funding shock after US withdrawal, making the leadership contest partly about institutional survival.
  • Chinese coverage highlights representation and continuity with China’s earlier support for Margaret Chan, while the source gives no competing candidates or election timetable.
  • A Chinese-led WHO could increase Beijing’s influence over global health priorities, but election arithmetic and the nominee’s policy record are not yet described.
Open at SCMP Ask Claude ChatGPT