China Desk — 11 Oct 2026, 05:35 UTC

ByteDance paper finds periodic blind spots in DeepSeek's long context

# Leiphone 17h

A ByteDance research paper finds that fixed-size KV-cache compression can make DeepSeek models highly sensitive to token position, causing harmless whitespace changes to alter answers.

In the reported code-completion test, adding decorative equals signs changed the answer; error rates reached 71.3 percent for some alignment phases versus 91.5 percent accuracy for others.
The mechanism is architectural: fixed-size chunks compress keys and values into position-dependent summaries, so tokens near weak slots or chunk boundaries can be underrepresented.
A 128,000-token retrieval test found maximum position-based accuracy gaps of 40.2 percent for one model and 34.8 percent for another; later post-training reduced but did not eliminate the effect.
Halving the compression stride improved the reported gap to 6.1 percent in a newer variant, while controlled smaller-model experiments suggested the periodic error is inherent to fixed chunking rather than a tuning or position-encoding bug.
The production risk is silent: contract clauses, code symbols, or safety instructions can fall into a compression blind spot without changing their semantic content.
The research points toward dynamic chunk boundaries or other content-aware compression, trading a clean memory-saving mechanism for a more complex runtime and hardware design.

Open at Leiphone Ask Claude ChatGPT

Xiaomi spends $2.6 million per RL training run on agentic scaling

# InfoQ China 12h

Xiaomi's MiMo-V2.6 report details a large-scale agentic reinforcement-learning system built around MoE models, million-token contexts, and expensive quality grading.

MiMo-V2.6 Pro has 1.02 trillion total parameters with 42 billion active per token, while Flash has 310 billion total and 15 billion active; both combine mixture-of-experts routing with local and global attention.
Each reinforcement-learning step samples 1,568 prompts with 16 rollouts each, producing about 25,000 trajectories and 2.7-3.7 billion tokens, with sequences averaging roughly 110,000-150,000 tokens.
The reported post-training costs are about $2.6 million for Pro and $900,000 for Flash; rollout and parameter training each account for roughly 44 percent of Pro cost, while grading consumes 12.7 percent.
Xiaomi's groupwise reward methods rank successful solutions by engineering quality, side effects, patch size, and verification rather than relying only on binary test pass rates.
The report also describes adversarial training environments that block leaked fixes and use a separate hack agent to find reward-hacking paths, showing that agentic RL is partly an evaluation and containment problem.
The strategic implication is that scaling agent capability now requires scaling environments, graders, and security controls alongside model parameters; compute alone will not produce reliable long-horizon behavior.

Open at InfoQ China Ask Claude ChatGPT

Zhejiang calls for limits on tech giants' control of online traffic

# SCMP 17h

Party officials in Zhejiang say platform traffic and algorithms now shape public discussion and sentiment, calling for stronger oversight of their power.

The provincial Party publicity department said online traffic should no longer be treated merely as a data resource, elevating algorithms and distribution systems into matters of social governance.
The language suggests a shift from regulating content and data separately toward regulating the ability of large platforms to set attention, visibility, and public mood.
Zhejiang matters because it is home to major internet companies and Hangzhou's platform economy; the statement is therefore both a governance signal and a warning to firms based there.
This is not yet a national rule, but provincial Party commentary can preview the direction of enforcement and the political vocabulary later used by central regulators.
For platform executives, the risk is that ranking, recommendation, and traffic allocation become treated as quasi-public powers subject to political accountability rather than ordinary product decisions.

Open at SCMP Ask Claude ChatGPT

Manus raises more than $500 million after its Meta break

# TMTPost 57m

Manus's Chinese parent raised more than $500 million at a reported valuation near $4 billion, reframing the startup from Meta target to independent agent platform.

The reported round was led by Boyu Capital and IDG Capital, with Tencent, Sequoia China, and ZhenFund participating; the source says the valuation was reported rather than officially disclosed.
The company's strategic reset followed Meta's canceled acquisition and a return to independent operation, with Manus pursuing overseas users while rebuilding a China-market product line.
Its Cue product gives agents separate email, phone, wallet, and cloud-computer identities and supports multi-agent collaboration, pushing beyond a single assistant toward an agent team.
The reported annual recurring revenue rose from more than $100 million at the end of 2025 to roughly $400-500 million by mid-2026, but the source does not provide audited figures or cohort economics.
Tencent's continued investment signals that Chinese investors see independent agent infrastructure as a scarce strategic asset, especially while major platforms have not settled on a winning agent form.
The central test is whether Manus can retain users once larger companies embed comparable capabilities into office suites, messaging, and operating systems.

Open at TMTPost Ask Claude ChatGPT

Huawei opens three native AI development tools for HarmonyOS PCs

# AI Zhidx 17h

Huawei put its desktop IDE, agentic coding environment, and command-line toolset into public beta, turning HarmonyOS PCs into native development workstations.

DevEco Studio provides the integrated IDE and multi-device debugging, DevEco Code adds goal-plan-build agent workflows, and DevEco CLI exposes more than 70 skills through command-line interfaces.
The architecture is designed for controlled autonomy: developers can approve a plan before execution or let the agent implement, compile, run, verify, and repair against a goal specification.
CLI packaging is strategically important because it makes HarmonyOS capabilities callable by external agents instead of trapping developers inside Huawei's own interface.
Huawei says HarmonyOS 6 and 7 devices have passed 90 million units, providing a potential installed base for an application ecosystem that needs better native tools.
The move is an ecosystem play: Huawei is trying to make its operating system not only a place where apps run but also a place where agents build, test, and ship them.

Open at Zhidx Ask Claude ChatGPT

WorldArena 2.0 tests whether world models work on robots

# AI Leiphone 1d

A Tsinghua-led benchmark expands world-model evaluation from video generation to touch, online reinforcement learning, and real-robot manipulation, with Chinese teams taking prominent places.

WorldArena 2.0 adds visual-tactile inputs, interactive online training, and physical robot evaluation to the earlier focus on videos, offline tasks, and simulation.
The benchmark is designed to separate visual realism from usefulness: a model that predicts attractive video may still fail to produce reliable rewards, action consequences, or policy improvement.
The challenge has three tracks for video quality, world models as reinforcement-learning environments, and real-robot operation; the first two reported winners came from an industrial JEPA team and an academic spatial-intelligence team.
This is a meaningful change in evaluation philosophy: embodied models are being judged on causal state prediction, compounding error, latency, and execution rather than on semantic plausibility alone.
The benchmark also exposes a strategic fault line between representation-first JEPA approaches and spatial-world-model systems, while forcing both to meet the same physical tasks.
The remaining limitation is scale and coverage: a small set of robot platforms and benchmark scenes cannot yet establish broad household or industrial reliability.

Open at Leiphone Ask Claude ChatGPT

Europe forces Tesla to soften its Full Self-Driving brand

# Policy QbitAI 22h

Chinese coverage frames Tesla's European rename as a regulatory concession made to secure approval, while also stressing that Chinese EVs are outgrowing Tesla in the region.

Tesla changed the European name to Tesla Assisted Driving after Germany said Full Self-Driving was misleading because the driver must remain attentive.
Germany also signaled support for European approval and accepted a negotiated speed deviation of up to 10 percent above the limit in defined conditions, showing the trade between naming discipline and market access.
The source says eight smaller European Union countries had approved the system, covering only about 12.6 percent of the bloc's population, while larger markets remained unresolved.
The Chinese framing connects the regulatory fight to competitive pressure: it reports BYD registrations above Tesla's in the same period and presents driver assistance as Tesla's main remaining differentiation.
The case shows how names, liability, and approval regimes can become part of product engineering strategy when a supervised system is marketed with autonomy language.

Open at QbitAI Ask Claude ChatGPT

Nio says AI-driven memory costs are creating diseconomies of scale

# Analysis Leiphone 21h

Nio's Li Bin says rising AI demand for memory and computing can make larger vehicle production more expensive, even as the company reports stronger margins and premium pricing.

Nio says second-quarter cost per vehicle rose 14,000 yuan, roughly $2,000, creating about 1.5 billion yuan, or $210 million, of quarterly cost pressure at its reported volume.
Despite that, the company reported first-half revenue of 57.67 billion yuan, roughly $8.1 billion, up 85.8 percent, and a second-quarter average transaction price of 406,000 yuan, roughly $57,000.
Li attributes the unusual cost curve to competition between automotive demand and AI data centers for memory and other scarce components, challenging the usual assumption that scale automatically lowers unit cost.
Nio's management is trying to convert its service, charging, and community infrastructure into a second profit engine, with nonvehicle revenue reportedly large enough to cover charging-network investment.
The company uses non-GAAP operating profit as evidence of improvement, but the source notes that this excludes stock compensation and leaves the durability of reported profitability open to debate.
The story captures a broader Chinese industrial tension: automakers are becoming AI-heavy systems companies while competing with data centers for the same hardware inputs.

Open at Leiphone Ask Claude ChatGPT

IROS 2026 shows embodied AI moving from hype to engineering

# AI Leiphone 1d

Chinese analysis of IROS 2026 sees embodied AI converging on physical world models, layered fast-and-slow control, tactile sensing, and cheaper data collection.

The reported research direction is shifting from language models that describe tasks toward models that internalize friction, gravity, deformation, uncertainty, and action consequences.
A recurring architecture separates a slow planning layer running language or vision models from a fast controller operating at roughly 1 kilohertz, allowing high-level reasoning without sacrificing physical stability.
The analysis highlights tactile sensing and sim-to-real transfer as central variables, suggesting that vision-only demonstrations are no longer an adequate proxy for manipulation skill.
One cited system reportedly runs a dynamic controller on a Raspberry Pi 4B and can update control policies without stopping the robot, illustrating the field's push toward practical systems rather than giant-model theater.
The Chinese framing is that academic and industrial boundaries are dissolving: model design, data collection, control software, and hardware deployment are being evaluated as one stack.
The broader signal is architectural convergence, not a single winning model; the hard work has moved to interfaces, latency budgets, calibration, and reliable physical feedback.

Open at Leiphone Ask Claude ChatGPT

Two forced tokens can unlock a base model's reasoning

# AI Zhidx 23h

Research from MIT and partner institutions finds that fixing the first two response tokens can lift a base model's math accuracy from 42 percent to 78 percent, near its reinforcement-trained version.

For Olmo-3-7B, forcing an opening equivalent to a brief acknowledgment raised MATH-500 accuracy from 42 percent to 78 percent, above the reported 75 percent result of its reinforcement-learning counterpart.
The authors argue that reinforcement learning may partly teach a model to select a productive reasoning trajectory at the start, rather than create all of the underlying capability from scratch.
The effect also appeared in a Qwen3-14B experiment, making the China dimension substantive even though the research institutions are international.
The finding is a prompt-control result, not a claim that two tokens replace reinforcement learning in general; transfer to long-form coding, tool use, or adversarial tasks remains unclear.
For model builders, it suggests that initial-state selection, decoding dynamics, and training-data associations deserve more attention as separate levers from parameter updates.

Open at Zhidx Ask Claude ChatGPT

Leapmotor promises an iPhone-style second car brand

# Analysis Leiphone 21h

Leapmotor says its second brand will redesign the cabin and vehicle computing stack around usable interior space, centralized control, and agent software rather than incremental styling.

Leapmotor says 8,500 of its 10,000 employees work in research and development and that it controls about 65 percent of vehicle cost through 18 component factories; these are company claims.
The proposed architecture centralizes vehicle control, cuts wiring below 500 meters, uses 48-volt power, and places long-term agent memory in the cloud while keeping latency-sensitive functions in the car.
Packaging changes target a cabin occupancy ratio above 100 percent through a combined HVAC unit, rear-mounted powertrain, flatter floor, and taller interior; the company claims a 4,400-millimeter cabin length and major space gains.
The strategy is an attempt to make the car a mobile room rather than a better version of the conventional five-seat layout, while also creating an external business from batteries, drives, and intelligence.
Leapmotor says its world-model driver assistance moved from route selection in late 2025 to a real-vehicle demonstration in under a year, but the source provides no independent validation of the claimed first-tier performance.
The hard question is whether this level of vertical integration can create a durable second brand or merely increase complexity in a market where software and hardware features are rapidly copied.

Open at Leiphone Ask Claude ChatGPT

Step 5 Preview tops OpenRouter's trending list

# AI Leiphone 1d

Chinese model maker StepFun says its 600-billion-parameter Step 5 Preview reached the top of OpenRouter's trending list and will open source it on October 15.

The model uses a sparse mixture-of-experts design with about 27 billion active parameters per token and is tuned for coding, software engineering, finance, and knowledge work.
The source reports a 44-point Artificial Analysis intelligence score and positions the model among the top open models, but the claims are company-linked and lack a full independent benchmark table.
OpenRouter rankings are usage signals rather than capability evaluations: trending rank measures recent growth, while daily rank reflects token volume, so both can reward price and availability as well as quality.
The strategic value is the combination of open weights, competitive cost, and a route into international developer workflows, giving a Chinese model company a way to compete outside domestic distribution channels.
The upcoming release will matter more than the preview ranking because reproducibility, license terms, inference cost, and performance under real agent workloads will determine whether developers keep it in their stack.

Open at Leiphone Ask Claude ChatGPT

Taiwan's leader welcomes US-China talks to reduce war risk

# Policy SCMP 19h

Taiwanese leader William Lai welcomed Washington-Beijing talks while warning that China is pressuring the island to alter its status.

Lai's address combines support for risk-reduction talks with a call for democracies to coordinate and for Taiwan's right to exist to be respected.
The dual message reflects Taiwan's attempt to support crisis management without accepting Beijing's political framing of the island's future.
For China watchers, the significance is the continued coupling of dialogue language with deterrence and internationalization rather than a move toward bilateral accommodation.
The report is a political signal, not evidence that talks have resolved the underlying dispute or reduced military pressure.

Open at SCMP Ask Claude ChatGPT

Nio narrows its next three years to premium focus

# Analysis Leiphone 21h

Nio's chief executive told employees that the next three years will focus on the core business, premium vehicles, and key markets while the company limits embodied-AI exposure to strategic investment.

Li Bin described China's auto market as entering its most brutal phase, yet said Nio's deliveries rose 57.9 percent year on year through August while the overall passenger-car market fell about 20 percent.
Nio says its three brands delivered 262,893 vehicles in the first eight months and that average prices remain above 400,000 yuan, roughly $56,000, for the flagship brand.
The company is treating charging, battery swapping, service, and community operations as a second growth curve built on vehicle ownership rather than as a standalone technology bet.
Li's decision to use investment and partnerships rather than make embodied intelligence a core business is a notable counterpoint to Chinese companies rushing into humanoids and robotics.
The speech is internal and the projections are management claims, but it reveals an attempt to replace expansion rhetoric with a narrower operating thesis: premium pricing, service monetization, and execution quality.

Open at Leiphone Ask Claude ChatGPT

Li Auto is turning in-house batteries into an industrial capability

# Analysis Leiphone 21h

Li Auto is moving self-developed 5C batteries across its lineup and taking an 8.79 percent stake in Sunwoda Power to control data, integration, and supply resilience.

The reported investment is 2.65 billion yuan, roughly $370 million, and would give Li Auto about 11.17 percent of Sunwoda Power through related entities.
Li Auto's strategy is not to manufacture every cell itself but to define materials, cells, packs, battery management, thermal systems, and vehicle integration as one product system.
The company argues that battery data must run from supplier manufacturing through vehicle validation and field monitoring so safety anomalies can be found earlier and responsibility remains with the automaker.
The move reflects a broader Chinese EV shift from buying batteries as standardized components toward treating them as a strategic software-and-hardware subsystem.
The key test is whether tighter integration produces measurable safety, charging, lifetime, and service advantages without saddling the automaker with the capital intensity and operational risk of a battery manufacturer.

Li Auto says battery ownership matters less than battery control Leiphone

Open at Leiphone Ask Claude ChatGPT

Chinese military aircraft may be using a joint airfield in Laos

# Policy SCMP 15h

Satellite imagery reportedly shows more trainer aircraft at a new Laotian facility than Laos is assessed to operate, raising questions about Chinese military access.

The imagery reportedly shows eight K-8 or JL-8-family trainer and light-attack aircraft at Ban Keun Airport, alongside more than two dozen military and support vehicles.
The discrepancy does not prove Chinese ownership or operational control, but it is a concrete indicator worth monitoring because aircraft basing can reveal security cooperation before formal announcements do.
A joint facility would extend China's military access and training footprint in mainland Southeast Asia without requiring a conventional overseas base narrative.
The story is based on imagery and reporting cited by the source, so aircraft identity, mission, and duration of presence remain uncertain.
The regional implication is strategic: Laos sits within China's continental security and transport network, and even limited aviation access could improve training, logistics, and influence.

Open at SCMP Ask Claude ChatGPT

China and Europe agree to curb Chinese hybrid-car exports

# Policy Jiemian 45m

Chinese coverage stresses a negotiated trade accommodation, while European reporting emphasizes a possible halving of Chinese hybrid-car exports and market share.

The reported understanding could cut Chinese hybrid and plug-in hybrid exports to Europe by several million vehicles over four years, roughly half the previously expected volume; the exact rules were not published.
The measure matters because European tariffs already hit Chinese battery-electric vehicles, pushing Chinese automakers toward hybrids as an export route that was not covered by the earlier subsidy investigation.
The source says Chinese brands hold about 29 percent of the Western European plug-in hybrid market and could be pushed toward 15 percent under the discussed approach.
Chinese manufacturers benefit from larger batteries, electric-drive-heavy designs, and a dense domestic supply chain, so the dispute is partly about whether European producers can compete on comparable cost terms.
For China, the implication is that export growth can no longer rely on volume alone; local production, supply-chain localization, and negotiated market access become necessary.
The agreement is strategically important but still incomplete: its enforcement mechanism, model-level allocation, and compatibility with World Trade Organization rules remain unclear.

Open at Jiemian Ask Claude ChatGPT

Hangzhou doubles down on deep tech without viral hype

# Analysis SCMP 4h

Hangzhou is promoting a quieter group of deep-tech bets in smart glasses, inference chips, compute-in-memory, and space computing after the city's earlier AI successes drew global attention.

The city's new portfolio is less social-media friendly than a robot demonstration or a popular open model, but it points to a deliberate attempt to build multiple layers of hardware and applied research capability.
The shift matters because Hangzhou's technology identity is moving from a few breakout companies toward an ecosystem of specialized firms that may need longer commercialization cycles.
The Chinese framing is notably more sober than hype-driven startup coverage: the question is whether local capital, talent, and industrial customers can support difficult hardware businesses after attention moves elsewhere.
This provides a useful view of how Chinese cities compete for technology status: through curated company portfolios, institutional backing, and industrial clustering as much as through individual product launches.

Open at SCMP Ask Claude ChatGPT

Luxeed wants to become Huawei's distinctive premium car brand

# Analysis Leiphone 21h

Luxeed says its next phase is to absorb Huawei's operating system and manufacturing discipline while building a product identity distinct from the other brands in Huawei's automotive portfolio.

The brand is trying to solve a structural problem created by Huawei's success: when multiple automakers share the same driving and cockpit technology, the partner brand can become interchangeable.
Luxeed says it has invested tens of billions of yuan, roughly several billion dollars, in a dedicated smart factory and is building its own research, design, sales, and service capabilities while retaining Huawei's ecosystem support.
The V9's reported 10,101 July deliveries at an average transaction price near 500,000 yuan, roughly $70,000, are presented as evidence that a Huawei-linked brand can command premium pricing.
The next RX model is positioned around driver-focused design and advanced sensing, not as a simple upgrade, suggesting the brand is trying to establish product personality within the larger Huawei system.
The strategic challenge is governance: Luxeed must gain enough independent organizational identity to differentiate itself without losing the technology, distribution, and operational leverage that make the Huawei partnership valuable.

Open at Leiphone Ask Claude ChatGPT

Pombot raises more than $30 million for AI sports hardware

# AI 36Kr 3h

Chinese sports-robotics company Pombot raised several hundred million yuan, roughly $30-60 million, to scale a vision-and-decision system that turns ball machines into AI coaches.

Pombot's architecture separates the physical serving system, high-speed visual sensing, and a sports model that converts observations into corrective training plans.
Its edge-cloud design puts low-latency perception and lightweight decisions on the device while using cloud models for richer analysis, then distills validated behavior back to the terminal.
The company reports more than 300,000 global users, over 2 billion balls served, and more than 1 million labeled ball trajectories, giving it closed-loop training data rather than generic sports video.
The product is designed around a causal loop: identify a technical problem, prescribe a drill, execute it with the machine, and measure whether the player's motion improves.
The strategic question is whether that data moat can transfer from table tennis and tennis to other sports without losing the sport-specific expertise that makes the system useful.

Open at 36Kr Ask Claude ChatGPT

Kuaishou is building an AI agent for mobile reliability

# AI InfoQ China 0m

Kuaishou's Conan agent targets the post-coding bottleneck: diagnosing crashes, freezes, and memory failures across Android, iOS, HarmonyOS, and shared-runtime stacks.

The presentation describes a production workflow that aggregates incident evidence, analyzes heap snapshots and cross-language failures, proposes remediation, and feeds fixes back into an observability loop.
Its architecture emphasizes workflow orchestration, context engineering, an agent harness, and data services rather than simply adding a chatbot to an existing monitoring product.
HarmonyOS and Kotlin Multiplatform increase the number of runtimes and compatibility histories the system must reconcile, making context retrieval and evidence quality as important as model capability.
The source explicitly says private operational knowledge and incomplete device-side evidence remain hard limits; the agent needs continuing infrastructure investment and controlled access to internal systems.
This is a useful Chinese example of agents moving from code generation into software operations, where the value is measured by closed-loop incident handling rather than token throughput.

Open at InfoQ China Ask Claude ChatGPT

A museum replica creates rare US-China collaboration

# Culture SCMP 7h

A life-size replica of a Yuan-era Buddhist mural at New York's Metropolitan Museum has enabled an unusual collaboration around preserving and displaying Chinese cultural heritage.

The project centers on a monumental Medicine Buddha assembly from the Yuan period, whose scale and fragility make ordinary exhibition and conservation difficult.
A full-size replica can let visitors experience the work while reducing pressure on the original, linking museum display, digital or physical reproduction, and conservation practice.
The story is valuable because it shows cultural exchange operating through technical craft and institutional trust rather than only through official diplomacy.
It also exposes a recurring tension in Chinese heritage: global access increases visibility, but reproducing and moving fragile works requires shared standards for authenticity, stewardship, and risk.

Open at SCMP Ask Claude ChatGPT

BYD rolls out its 150,000th electric commercial vehicle

# Analysis Leiphone 1d

BYD's 150,000th new-energy commercial vehicle is a 600-kilowatt electric tractor, signaling a shift from buses toward long-haul truck electrification.

The Q3 tractor uses BYD's own three-electric-system technology, a commercial blade battery, and dual-gun fast charging for heavy freight applications.
The milestone matters because long-haul trucks are a harder electrification problem than city buses: they require high uptime, payload performance, fast replenishment, and dependable service networks.
BYD says its commercial vehicles now operate in more than 70 countries and is pitching an integrated package of vehicles, charging, fleet management, and lifecycle maintenance.
The company is therefore moving from selling electric vehicles to exporting an operating system for zero-emission logistics, although the source provides no independent data on total cost or fleet uptime.

Open at Leiphone Ask Claude ChatGPT

China warns that a few seconds of speech can enable voice cloning

# Policy IT Home 2h

China's Ministry of State Security warns that short voice samples, facial images, and other biometrics can feed deepfake fraud and targeted intrusion.

The ministry says an attacker can bait a person into speaking during a silent phone call and use the resulting sample for voice-cloning scams; this is an official warning, not a measured estimate of attack success.
The broader message treats voice, face, iris, and gait data as durable identifiers that can be copied, combined, and inferred across databases.
The warning links biometric collection to both consumer fraud and national-security threats, giving privacy protection a security framing rather than treating it only as a data-governance issue.
It also reinforces China's principle of collecting biometric information only when necessary and warns that illegal sale or distribution of facial data can constitute a criminal offense.
The practical implication for companies is that deepfake risk is shifting from model quality to data provenance, consent, access control, and the ability to prove that a biometric signal was not synthesized.

Open at IT Home Ask Claude ChatGPT

China proposes stricter safety checks for novel car controls

# Policy IT Home 2h

China's industry regulators propose front-loading safety analysis and banning fully hidden door handles, while restricting risky seat and display designs.

The draft guidance would require failure-mode analysis under complex and extreme conditions before novel controls are installed in vehicles.
It calls for physical, clearly marked emergency controls, mechanical redundancy and hand clearance for door handles, and a ban on fully hidden handles.
The proposal would also restrict zero-gravity, rotating, or bed-style seats while driving and require safety evaluation of large screens, head-up displays, and complex human-machine interfaces.
This is a state response to a Chinese auto market that rewards visible novelty but increasingly exposes the safety cost of treating interface design as a styling contest.
If adopted, the rules would move responsibility from post-launch recalls toward design-stage evidence and could slow the rapid spread of unproven cabin features.

Open at IT Home Ask Claude ChatGPT

The AI jobs being created may not be open to ordinary workers

# Culture TMTPost 27m

China's expanding AI job titles expose a harder problem: automation may remove entry-level training ladders before displaced workers can qualify for the new roles.

China's labor authorities added 11 occupations and 23 job categories, including agent development, embodied-robot applications, and digital-twin engineering, but a new label does not create an accessible career path.
The article contrasts high-skill AI roles with lower-paid robot data collection and review jobs, showing how AI can create employment while widening the gap between new work and desirable work.
The deeper risk is loss of apprenticeship: coding agents can absorb the debugging and basic implementation tasks through which junior engineers traditionally built judgment.
The cited research is not China-specific, but the issue matters locally because China's large graduate and service-sector labor pools face the same transition bottleneck.
The key policy question is whether training, hiring, and enterprise organization create an on-ramp, rather than merely publishing a more futuristic occupational catalog.

Open at TMTPost Ask Claude ChatGPT

China's dancing robots are becoming a tool of African outreach

# Culture SCMP 21h

Chinese embassies and companies used dancing robots and robot dogs at National Day events in Africa, turning consumer robotics into a soft-power and business signal.

The demonstrations in Angola and Congo presented robots as visible symbols of Chinese industrial upgrading rather than as narrowly technical products.
The events put Chinese officials, local government representatives, and business communities in the same room, linking entertainment, diplomacy, and commercial relationship-building.
The use of inexpensive, legible demonstrations reflects a broader Chinese export strategy: show embodied technology in public before asking partners to understand the underlying supply chain or industrial system.
The robots may be promotional, but their value lies in making China's manufacturing and AI narrative tangible to audiences that do not encounter Chinese research papers or factory automation directly.

Open at SCMP Ask Claude ChatGPT

3D chip stacking faces a yield and cost problem

# Analysis TMTPost 1h

A Chinese technical analysis explains why 3D integration can improve performance while making each good chip more expensive through compounding yield loss and difficult bonding.

The article models two 95 percent-yielding layers producing only a 90.25 percent combined yield before alignment, bonding, testing, and later failures are counted.
It contrasts wafer-to-wafer bonding, which is efficient but lets bad dies contaminate good pairings, with die-to-wafer bonding, which costs more per placement but screens out bad dies first.
Under the article's illustrative assumptions, die-to-wafer processing produces a lower cost per passing stack, about $99 versus $108, despite a higher bonding charge; the figures are explicitly hypothetical.
The practical bottleneck is the whole process chain: cleaning, planarization, alignment, bonding, annealing, inspection, and feedback, not just the bonding machine.
The analysis places Chinese and foreign companies on the same technology map, including Huawei, but does not establish a China-specific manufacturing breakthrough.

Open at TMTPost Ask Claude ChatGPT

China's IELTS exams were canceled nationwide after technical failures

# Culture IT Home 2h

Widespread IELTS computer-test cancellations in China disrupted applicants during admissions season and exposed the fragility of a forced shift to fully digital exams.

The October 10 sessions were canceled across much of mainland China, with some candidates stopped after the listening section; the organizer said fees would be returned but offered no reported compensation for travel or lodging.
The timing is unusually damaging because applicants to universities in Hong Kong, Britain, and Singapore often need scores during the September-to-October application window.
China ended ordinary paper IELTS tests in September 2026, so the incident is a stress test of nationwide dependence on computer delivery rather than an isolated inconvenience.
The episode shows how digitization transfers operational risk to candidates: a platform outage can invalidate a test day for thousands while leaving individuals to absorb deadline and travel consequences.
The official explanation was technical failure, but the wider lesson is about resilience, contingency capacity, and accountability when high-stakes services are centralized.

Open at IT Home Ask Claude ChatGPT

AI glasses still look like a compromise product

# Culture TMTPost 51m

Chinese consumer analysis finds AI glasses selling mainly as camera-and-audio devices, not as a convincing replacement for the smartphone.

The cited market data says Chinese smart-glasses shipments reached 2.46 million units in 2025, up 87.1 percent, while screenless products made up about 96 percent of global AI-glasses shipments in the first half of 2026.
That split suggests early demand is for first-person capture, open-ear audio, and voice control; visual displays and persistent assistant behavior remain secondary.
The article argues that many AI-glasses functions can already be delivered by phones or earbuds, so the category lacks a frequent, hard-to-substitute use case for ordinary consumers.
China's market is useful here because rapid shipment growth coexists with uncertainty about why people would keep wearing the devices after the novelty fades.
The strategic battle is over a future interface and data-collection position, but vendor interest in a new entry point does not prove consumer demand for another device.

Open at TMTPost Ask Claude ChatGPT

Meitu's AI revival is real, but its stock remains volatile

# Analysis TMTPost 1h

Meitu's shift from a consumer beautification app to an AI imaging and design subscription business has restored growth, even as its share price has fallen by more than half from its 2026 peak.

Meitu reported first-half revenue of 2.213 billion yuan, roughly $312 million, up 22.1 percent, with adjusted profit of 652 million yuan, roughly $92 million, up 39.5 percent.
Its imaging and design business grew from 1.33 billion yuan in 2023 to 2.95 billion yuan in 2025, reaching 1.769 billion yuan in the first half of 2026.
The company embedded its MiracleVision visual model across its product suite and shifted monetization toward subscriptions, SaaS-like creative tools, and professional users.
It reported more than 18.44 million paid subscribers and 33 million monthly active users for productivity applications, illustrating how AI can improve monetization of an established consumer tool without turning the company into a foundation-model vendor.
The stock's volatility shows the difference between a credible operating turnaround and a stable market narrative: AI expansion has improved the business, but investors still price it as a consumer software company exposed to rapidly changing product fashions.

Open at TMTPost Ask Claude ChatGPT

Avatr chairman shifts to Changan's overseas business

# Analysis Leiphone 21h

Avatr chairman Wang Hui is reportedly moving to lead Changan's Southeast Asian and Latin American expansion while Avatr president Chen Zhuo takes on more operating responsibility.

The change would separate Avatr's IPO-push role from day-to-day operating control while keeping the brand within Changan's broader international strategy.
Wang had also been an early advocate of Changan's Huawei partnership, so the move may reflect a redistribution of executive attention rather than a retreat from the alliance.
The report is not formally confirmed: internal systems still list Wang as chairman and the source describes an adjustment in progress.

Open at Leiphone Ask Claude ChatGPT

DJI and Insta360 dominate the smart-camera market

# Signal Techmeme 10h

US coverage cites IDC estimates giving Shenzhen's DJI and Insta360 73 percent and 20 percent of the global smart-camera market in the second quarter.

The combined share attributed to the two Chinese companies is about 93 percent, while GoPro has fallen to about 3 percent, according to the cited IDC figures.
This is more than a product story: Chinese firms now control the leading action-camera and smart-camera platforms through hardware integration, imaging software, and global distribution.
The rivalry between DJI and Insta360 suggests the next competitive contest is not China versus a Western incumbent but two Chinese companies fighting over the highest-value global category position.
Because the item comes through US financial press and an aggregator, the market-share figures should be checked against the underlying IDC release before making investment decisions.

Open at Techmeme Ask Claude ChatGPT

Two Qing-era ceramics damaged at the British Museum

# Culture SCMP 18h

Chinese ceramics were damaged during display at the British Museum, renewing debate over how overseas institutions protect fragile Chinese cultural objects.

The museum says a bowl fell from its shelf onto a small box, damaging both objects; conservators are restoring them.
The incident matters beyond the physical damage because Chinese cultural objects abroad sit at the intersection of heritage access, museum accountability, and national sensitivity.
Chinese coverage frames the event as another test of Western museum stewardship, while the museum's response emphasizes accidental damage and professional conservation.
The source does not establish whether the display violated a specific standard, so broader claims about systemic security should remain tentative.

Open at SCMP Ask Claude ChatGPT

ChinAI	403	0/0		HTTPError: 403 Client Error: Forbidden for url: https://chinai.substack.com/feed
ChinaTalk	200	0/20
Pekingnology	200	0/20
Ginger River Review	200	0/20
Sinocism	200	0/20
QbitAI	200	1/10
36Kr	200	2/30
Solidot	200	5/20
Leiphone	200	15/20
GeekPark	200	3/30
IT Home	200	12/60	+39 over cap
SCMP	200	9/50
Initium	200	1/15
Jiemian	200	8/30	+22 over cap
China Digital Times	200	0/15
The Wire China	200	0/50
Techmeme	200	4/15
Hacker News	200	0/20
V2EX	200	8/50	+13 over cap
Zhidx	200	4/20
InfoQ China	200	8/20	+10 over cap
New Wisdom	200	8/15	+2 over cap
TMTPost	200	8/20	+12 over cap

The cheapest part on a million-yuan car is about to move a national standard

A state lab starts asking about plastic

Zhejiang says traffic is a power now

Taipei welcomes the talks, Shanghai warns about them

Xiaomi published the bill for one RL run

The travel agent that sent you the wrong way

Threads we are pulling