China Desk — 10 Oct 2026, 14:25 UTC

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

# Leiphone 2h

A ByteDance paper argues that fixed-size KV-cache compression creates position-dependent recall failures: adding harmless spaces can shift key information into a periodic blind spot.

In the reported code-completion test, changing the count of decorative equals signs produced error rates of 71.3% at some phases and 91.5% accuracy at others.
The proposed mechanism is structural: fixed-size chunks compress tokens into learned summaries, making a token's position within a chunk matter in a way full attention does not.
On a 128,000-token needle-in-a-haystack test, accuracy varied by as much as 40.2 percentage points across positions for one DeepSeek base model and 34.8 points for another.
Halving the compression stride reduced the gap to 6.1% in the reported newer model, but did not eliminate the effect.
Controlled Qwen experiments found the periodic behavior whenever fixed-length compression was used, including with simple averaging and without rotary position encoding.
For production systems, the danger is silent: the same document and question can produce different answers after harmless formatting changes, making conventional regression tests insufficient.
Dynamic chunk boundaries are suggested as a possible remedy, but the paper's main contribution is to show that cheaper long-context memory can create systematic, phase-locked failures.

Open at Leiphone Ask Claude ChatGPT

WorldArena 2.0 tests whether world models work on robots

# Leiphone 8h

A Tsinghua-led benchmark moves world-model evaluation from video quality toward tactile input, online reinforcement learning, and real-robot execution, with Chinese teams prominent in the results.

WorldArena 2.0 expands along three axes: visual input becomes vision plus touch, offline tasks become online policy interaction, and simulation is supplemented by real robots.
Its three tracks separately test video quality, use as a reinforcement-learning environment, and real-robot manipulation, exposing the gap between plausible frames and useful dynamics.
The benchmark reports a JEPA-based team winning the video track and a spatial-intelligence model from Tongji University close behind, while another JEPA team placed second in the online-learning track.
The design matters architecturally: an error in predicted state can compound over many interactions, so a world model must provide stable transitions and useful reward signals, not just realistic pixels.
Putting real execution into a major robotics conference creates a common test for academic models and industrial systems that previously used incompatible demonstrations.
The remaining limitation is scope: the benchmark still covers selected tasks and robot platforms, so it cannot yet establish general-purpose physical intelligence.
For engineers, the key shift is evaluation of model utility inside a control loop rather than ranking generative video in isolation.

Open at Leiphone Ask Claude ChatGPT

Xiaomi details a $2.6 million scaling run for agentic reinforcement learning

# InfoQ China 0m

Xiaomi's MiMo report exposes the infrastructure economics of agentic RL: each step generates 2.7-3.7 billion tokens, with graders consuming a material share of post-training cost.

MiMo V2.6 Pro uses a 1.02 trillion-parameter sparse mixture-of-experts model with 42 billion active parameters; Flash uses 310 billion total and 15 billion active parameters.
Each reinforcement-learning step samples 1,568 prompts and 16 rollouts per prompt, producing about 25,000 trajectories and contexts up to one million tokens.
The training system separates rollout, grading, and parameter updates so each stage can scale independently; this is closer to a distributed production pipeline than a single training job.
Pro post-training cost about $2.6 million and Flash about $900,000; Pro spending was split roughly 43.8% rollout, 43.5% training, and 12.7% grading.
Xiaomi's groupwise graders compare successful patches on engineering quality, side effects, scope, and verification rather than rewarding tests alone.
The report also describes network isolation and adversarial agents to stop code agents from retrieving upstream fixes, illustrating that reward hacking is now a systems-security problem.
The route matters if these costs buy robust long-horizon behavior; otherwise, longer context and more graders may simply make failure more expensive.

Open at InfoQ China Ask Claude ChatGPT

Zhejiang Party officials call for tighter oversight of platform traffic

# SCMP 2h

Zhejiang's Party publicity department says platform traffic and algorithms now shape public sentiment, framing traffic power as a governance issue rather than a neutral data resource.

The commentary treats traffic allocation, recommendation systems, and platform scale as forces capable of influencing public discussion and social mood.
That language extends China's platform-regulation agenda from data security and competition toward control over attention and agenda formation.
Zhejiang is a consequential setting because Hangzhou hosts several major internet companies; local officials are signaling that platform governance is also political governance.
The framing does not announce a specific new rule, but it creates policy justification for stronger oversight of recommendation systems and traffic distribution.
For companies, the strategic risk is that engagement optimization can be judged not only by market or safety metrics but by its perceived effect on public sentiment.
The missing question is institutional: whether this remains commentary or becomes a concrete compliance, reporting, or algorithm-audit regime.

Open at SCMP Ask Claude ChatGPT

Analysts doubt Taiwan will derail the latest US-China reset

# SCMP 11h

Chinese and regional analysis treats Taiwan as a manageable complication in the current Washington-Beijing thaw, while warning that a reported $14 billion arms package remains a major fault line.

The article says Beijing is taking a pragmatic approach to the Taiwan issue while trying to preserve momentum in broader US-China relations.
The reported arms package matters because it can convert diplomatic tension into a concrete military and political commitment even when leaders are pursuing dialogue.
The strategic question is whether both governments can compartmentalize Taiwan-related actions or whether domestic political pressure will make every arms decision a test of credibility.
A temporary reset can reduce miscalculation without resolving sovereignty, alliance commitments, or the military balance around the island.
For technical companies, the relevant risk is not only conflict but the sudden return of export controls, investment restrictions, and logistics disruption after a period of commercial normalization.

Open at SCMP Ask Claude ChatGPT

Taiwan leader backs US-China talks to lower war risk

# SCMP 4h

Taiwan's leader welcomes Washington-Beijing talks to reduce conflict risk, while stressing democratic coordination and opposition to pressure from mainland China.

The address combines support for dialogue with a call for Taiwan's right to exist to be respected, keeping deterrence and political identity in the same frame.
Chinese coverage of the speech is likely to foreground the warning about mainland pressure, while Western framing emphasizes Lai's support for US-China risk reduction.
The episode shows how Taiwan's leadership tries to avoid being cast as the obstacle to détente while rejecting Beijing's preferred political terms.
For technology and investment planning, the important signal is that cross-strait risk management remains separate from resolution of the underlying sovereignty dispute.
The practical variable is whether Washington-Beijing engagement produces guardrails around military activity or only temporary diplomatic language.

Open at SCMP Ask Claude ChatGPT

Chinese military aircraft may be using a new Laos air facility

# SCMP 25m

Satellite imagery shows eight Chinese-built trainer or light-attack aircraft at a Laotian base, twice the assessed Lao inventory, raising questions about a joint operating site.

The imagery reportedly shows eight K-8 or JL-8 family aircraft and more than two dozen military and support vehicles at Ban Keun Airport.
The mismatch between visible aircraft and Laos's assessed inventory is suggestive, not proof that Chinese forces operate the facility; the source uses cautious language.
If confirmed, the site would give Beijing another practical layer of defense cooperation in mainland Southeast Asia without requiring a formal overseas base.
The development fits China's broader pattern of exporting equipment, training, and operating relationships rather than announcing a conventional permanent deployment.
The key missing evidence is aircraft ownership, crew identity, flight activity, and the legal scope of the alleged joint arrangement.
Commercial satellite imagery can establish presence but not intent, so the strategic conclusion should remain conditional.

Open at SCMP Ask Claude ChatGPT

Huawei opens three native AI development tools for HarmonyOS PCs

# AI Zhidx 2h

Huawei is beta-testing an IDE, an agentic coding tool, and a command-line layer for HarmonyOS PCs, turning the device from an app endpoint into a domestic development workstation.

DevEco Studio provides the IDE and multi-device debugging layer, while DevEco Code handles goal, plan, build, compilation, execution, and verification workflows.
DevEco CLI exposes more than 70 skills and HarmonyOS tools through command-line interfaces that external agents can call, reducing the need for bespoke integration scripts.
The architecture separates approval-heavy Plan and Build flows from a more autonomous Goal mode, reflecting Chinese vendors' attempt to make agentic coding controllable rather than merely generative.
Huawei says HarmonyOS 6 and 7 devices have surpassed 90 million, creating a domestic installed base large enough to justify a native toolchain.
The strategic objective is ecosystem lock-in: developers can build, test, and publish inside the same operating-system environment instead of relying on foreign desktop tooling.
The beta remains an ecosystem signal rather than proof of developer adoption; the crucial metrics will be tool stability, API coverage, and whether agents can complete real release workflows.

Open at Zhidx Ask Claude ChatGPT

Li Auto is turning battery research into an industrial capability

# Analysis Leiphone 5h

Li Auto is moving its internally developed 5C batteries across its lineup and taking an 8.79% stake in Sunwoda Power, using vertical integration to control vehicle-level performance and risk.

Li Auto plans to invest 2.65 billion yuan, roughly $373 million, in Sunwoda Power; including an earlier holding, its related entities would own about 11.17%.
The company says its battery work spans cell chemistry, pack design, battery management, thermal systems, and vehicle integration rather than simply making a standard component.
The move reflects the particular demands of China's range-extender market, where batteries must support high power, frequent cycling, and engine intervention at the same time.
A strategic partnership with a cell manufacturer lets Li Auto define the product and data requirements without owning every manufacturing asset.
The company is extending internally developed batteries from L8, L6, and i8 models toward the next Mega, i9, and i6, turning a technology program into a supply-chain commitment.
The risk is execution: five-C charging requires coordinated materials, cooling, safety, life, and software controls, and volume expansion will expose any weak link.

Li Auto says its batteries are about data control, not supplier replacement Leiphone

Open at Leiphone Ask Claude ChatGPT

Nio says AI is making some auto inputs more expensive at scale

# Analysis Leiphone 6h

Nio's Li Bin argues that AI demand for memory and data-center capacity is creating diseconomies of scale for automakers even as Nio reports stronger margins and sustained operating profit.

Nio says per-vehicle cost rose 14,000 yuan, roughly $2,000, from the end of last year, adding about 1.5 billion yuan, roughly $211 million, to one quarter's cost base at its reported volume.
The company reported first-half deliveries up 67.4% to 191,100, revenue up 85.8% to 57.67 billion yuan, roughly $8.1 billion, and a gross margin of 18.7%.
Li links the cost pressure to competition with AI data centers for memory and other scarce resources, reversing the traditional assumption that volume automatically lowers unit cost.
Nio's reported operating profit is on a non-GAAP basis, so the quality and durability of the turnaround remain contested even if the operating trend is real.
The company is using AI internally to improve forecasting and operating efficiency while keeping embodied intelligence as a partnership and investment area rather than a core business.
The broader signal is that AI infrastructure is becoming a cross-industry input shock: automotive scale can increase exposure to constrained compute components instead of reducing it.

Open at Leiphone Ask Claude ChatGPT

The V800 brake failure exposes a premium-car safety gap

# Policy 36Kr 14h

A repeated brake-pedal failure in a million-yuan car has triggered an industry standards review, while Chinese coverage questions whether low-cost mechanical parts receive enough engineering investment.

Also: IT Home

The reported test showed three vehicles breaking at the same plastic reinforcement around the pedal pivot during repeated emergency stops.
The automaker says its system exceeded national requirements and that no comparable base failure has appeared in ordinary customer use, but it plans a design improvement for more safety margin.
Chinese rules set limited force and deformation tests, leaving much of the real load requirement to each manufacturer's internal standard.
The article argues that premium branding and advanced sensors cannot compensate for weak validation of a low-cost mechanical load path.
Traditional luxury brands are presented as using stricter internal tests across temperatures and side-loading conditions, though those comparisons are industry claims in the source.
The regulatory significance is broader than this model: the industry may have to formalize material, load, fatigue, and supplier-development requirements for brake assemblies.

China may revise brake-pedal standards after a luxury-car failure IT Home

Open at 36Kr Ask Claude ChatGPT

China's inference boom is reshaping the SSD hierarchy

# Signal Leiphone 12h

At a Chinese storage conference, speakers said domestic token use has reached 140 trillion per day, pushing SSDs toward differentiated roles for KV cache, retrieval, and persistent model data.

The reported 140 trillion daily tokens are a conference claim, but they capture the scale shift from occasional model training to continuously served inference.
As agents retain long contexts, user sessions, and runtime state, storage must handle a mixture of low-latency reads, high throughput, large capacity, and endurance rather than one generic workload.
The emerging hierarchy places directly attached SSDs near GPUs for hot KV data, networked SSDs at larger scale, and capacity-oriented devices farther from compute.
Storage vendors are responding with workload-aware quality-of-service controls so one drive can allocate resources differently during planning, retrieval, and execution.
The article also describes QLC SSDs replacing some hard drives in data-center capacity roles and new flash designs aimed at model-weight and inference access patterns.
The strategic consequence is that AI infrastructure spending is broadening beyond HBM: storage architecture becomes part of model-serving performance and cost.

Open at Leiphone Ask Claude ChatGPT

Step 5 Preview tops OpenRouter's trending chart before open release

# AI Leiphone 10h

Chinese model maker StepFun's Step 5 Preview is gaining overseas developer usage and is scheduled for open release, showing how domestic labs are competing through accessible cost-performance.

The model reportedly reached first place on OpenRouter's trending list and second on its daily token-volume ranking, which measure usage momentum rather than benchmark quality alone.
It uses a 600-billion-parameter sparse mixture-of-experts architecture with about 27 billion active parameters per token and is optimized for coding, software engineering, finance, and research.
The source cites a composite intelligence score of 44 and a top-three open-model position, but these are external benchmark claims rather than independent production evidence.
The global distribution channel matters: an OpenRouter listing lets Chinese model developers reach users without first building their own international application layer.
The strategic contest is moving from model release announcements to actual token consumption, price, latency, and compatibility inside developers' existing agents.
The planned open release will be the real test of whether the preview's ranking reflects durable demand or short-lived curiosity.

Open at Leiphone Ask Claude ChatGPT

Xinjiang's desert seafood boom carries hidden costs

# Analysis The Wire China 1d

A Chinese-focused investigation says Xinjiang's fish-farming boom is presented as technology overcoming geography, while political control, imported feed, and environmental burdens complicate the story.

The central narrative is state-led transformation: aquaculture in a desert is used to demonstrate that engineering and policy can overcome natural constraints.
The investigation argues that the apparent food-production miracle depends on imported inputs and administrative control, so local output does not necessarily mean a self-sufficient system.
The story matters because Xinjiang development campaigns often combine economic targets, infrastructure, and political messaging; agricultural success is part of that governance model.
Its environmental and resource costs are the missing half of official success metrics, especially where water, feed, and logistics determine whether the model scales.
The contrast between promotional language and operational burden is more important than any single farm: it is a recurring pattern in China's regional development narratives.

Open at The Wire China Ask Claude ChatGPT

Nio tells employees to focus on premium cars and core markets

# Analysis Leiphone 6h

Li Bin says Nio will focus on its core vehicles, premium positioning, and key markets while avoiding a direct move into embodied AI as a main business.

Li says China's passenger-car market fell more than 20% year to date while Nio deliveries rose 57.9% through August, framing the company as growing against a contracting market.
The three-year plan is to focus on the core business, high-end products, and core markets rather than chase every technology fashion.
Nio plans to explore embodied intelligence through strategic investment and partnerships, not by turning robotics into an operating division.
The company says services and community revenue exceeded 10 billion yuan, roughly $1.4 billion, in 2025 and can now cover charging and battery-swap infrastructure costs.
The speech presents recurring services and infrastructure as a second growth curve built on installed base, not as a speculative new product category.
The strategic test is whether premium pricing, service economics, and pure-electric adoption can sustain growth while China's auto market remains brutally competitive.

Open at Leiphone Ask Claude ChatGPT

China's 2.1 billion holiday trips were mediated by platforms

# Culture TMTPost 5h

National Day travel data shows Chinese platforms managing navigation, pickup points, food discovery, and bookings at enormous scale, while also shaping where visitors go and what they buy.

China recorded 2.142 billion cross-regional trips from October 1 to 7, with nearly 80% by road; the scale turns maps and recommendation systems into public infrastructure.
Gaode reported nearly 370 million daily users, 9.7 billion kilometers of driving navigation, and more than 2.8 billion spatial-intelligence services on the first day.
Its local discovery list drew more than 128 million users, while Meituan's local recommendation traffic rose sharply in smaller cities, extending platform influence beyond major tourist centers.
Ride-hailing companies built designated pickup zones at hundreds of airports and stations, reducing the need for travelers to understand unfamiliar urban transport layouts.
The strategic shift is from search to delegated decisions: platforms increasingly decide routes, stops, restaurants, and transactions in one flow.
This convenience also concentrates cultural and commercial power in ranking systems whose data sources and incentives are largely opaque.

China's holiday travelers outsource more planning to AI TMTPost

Open at TMTPost Ask Claude ChatGPT

Chinese AI and hardware markets face a cost-and-control squeeze

# Analysis GeekPark 12h

A Chinese weekly roundup links memory inflation, Huawei licensing, AI-agent safety disputes, and domestic model distribution, showing how cost and governance are converging.

The roundup says a mainstream laptop's CPU, memory, and SSD account for about 68% of bill-of-materials cost, leaving brands to raise prices, cut margins, or reduce configurations.
It reports Huawei smartphone price increases and broader component-cost pressure, connecting the AI server buildout to consumer electronics rather than treating them as separate markets.
The China-specific AI items include Zhipu model distribution through overseas cloud revenue sharing and ByteDance's merger of office and coding agent products into one platform.
The foreign safety stories are included as context for Chinese readers, but their value here is the domestic contrast: model commercialization is being discussed alongside access, governance, and cost control.
The roundup's strongest signal is that Chinese AI firms are building distribution and monetization channels while hardware scarcity raises the cost of every deployment.

Open at GeekPark Ask Claude ChatGPT

A provincial Party boss meets Chinese creators over coffee

# Policy SCMP 12h

Jilin's Party secretary invited digital creators to a coffee shop and urged them to spread mainstream values, showing how propaganda work is adapting to influencer culture.

The informal setting is deliberate: local officials are treating creators as intermediaries who can shape public sentiment beyond traditional state media channels.
The official language of mainstream values and positive energy places online creativity inside the Party's broader content-governance framework.
This is not simply a publicity event; it is an effort to recruit platform-native voices into the state's preferred narrative system.
The episode illustrates how Chinese political communication is moving from command-and-control media toward managed participation by local creators.

Open at SCMP Ask Claude ChatGPT

Chery and Huawei want Luxeed to become a standalone premium brand

# Analysis Leiphone 6h

Luxeed says it wants to absorb Huawei's operating system and technology while building its own manufacturing, channel, and service identity inside the wider Harmony Auto ecosystem.

The brand says it has more than 6,000 employees, dedicated research and manufacturing facilities, and a growing independent sales network while still using Huawei's retail footprint.
Its Wuhu factory combines Huawei supply-chain management with AI quality monitoring, showing that the partnership is extending from cockpit and driving technology into industrial operations.
The strategic problem is differentiation: as more Chinese brands share Huawei technology, the partner brand must create its own product character rather than becoming a shelf label.
Luxeed's V9 premium MPV reportedly reached an average transaction price near 500,000 yuan, roughly $70,000, suggesting that the partnership can support pricing above the mass market.
The RX SUV is being positioned around driver-oriented handling and design, not merely as another Huawei-enabled family vehicle.
This is an organizational experiment in shared capability and independent brand equity, with manufacturing and after-sales closure still incomplete.

Open at Leiphone Ask Claude ChatGPT

Leapmotor says its second brand will remake the car around living space

# Analysis Leiphone 6h

Leapmotor's proposed second brand uses centralized electronics, a rear powertrain, and cloud-edge agents to turn cabin space and software architecture into the product's main differentiator.

The company says 8,500 of its 10,000 employees work in research and development, while it controls about 65% of vehicle cost through in-house components.
Its planned architecture puts a central domain controller in charge, uses 48-volt power, compresses wiring below 500 meters, and separates cloud reasoning from low-latency vehicle control.
A highly integrated HVAC system and rear-mounted powertrain are intended to increase cabin length, width, and height rather than merely add features to a conventional layout.
Leapmotor says its world-model driving stack moved from route selection in 2025 to a road-going demonstration in 2026, but the source supplies no independent performance evidence.
The second brand is planned for late 2027, so the strategic claims are an architectural direction, not a product currently available to test.
The interesting bet is that autonomous driving will make the cabin a usable living and work space, changing vehicle form factor more than propulsion alone did.

Open at Leiphone Ask Claude ChatGPT

Luxshare faces a new USITC patent investigation

# Policy Jiemian 1h

A US trade investigation names Luxshare and its subsidiary over a patent dispute, while Chinese coverage stresses that the case is early and the products are not yet in mass production.

The US International Trade Commission opened investigation 337-TA-1526 after naming Luxshare Precision and Dongguan Luxshare Technology as respondents.
The source says the products are still in customer validation and have not entered mass production, limiting immediate supply-chain impact.
Chinese coverage emphasizes procedural uncertainty and the absence of a substantive infringement finding, a familiar defensive frame in cross-border intellectual-property disputes.
For Luxshare, the strategic risk is not only an exclusion order but the possibility that customers delay validation or diversify suppliers while the case proceeds.
The episode shows how Chinese component companies can become exposed to US trade remedies before a product reaches commercial scale.

Open at Jiemian Ask Claude ChatGPT

Chinese enterprises are starting to manage AI by task cost

# Analysis TMTPost 4h

Chinese business coverage says token budgets are being cut as agents scale, exposing a management gap between visible AI spend and hard-to-measure business value.

The article cites surveys in which 61% of companies review AI cost during approval and 59% monitor it during operations, while only 12% continuously compare cost with value.
It says Chinese production deployments rose to 64.4%, but 60.4% of surveyed firms still lack a systematic value framework and 44.4% cannot clearly calculate AI project return.
Agent workloads change the cost model: spending follows business traffic and long-running tool use rather than employee chat frequency.
The proposed response is a harness layer that combines permissions, orchestration, accountability, and economic controls from the start.
The Chinese framing is practical rather than model-centric: the bottleneck is organizational redesign and unit economics, not access to a more capable model.
For executives, the useful metric is cost per successfully completed business task, with permission and failure budgets treated as first-class controls.

Open at TMTPost Ask Claude ChatGPT

China's tech and consumer markets show both AI adoption and pushback

# Culture 36Kr 14h

A Chinese daily roundup combines AI agents for household services, phone price rises, car-safety backlash, and growing concern about privacy and platform control.

ByteDance's Doubao is adding utility payments and travel services, moving from conversational answers toward account-linked transactions in daily life.
The roundup also reports smartphone price increases as memory costs rise, showing how AI infrastructure inflation is reaching ordinary Chinese consumers.
A reported brake failure in a premium vehicle triggered market punishment and regulatory attention, reinforcing that Chinese buyers still judge high-tech products through basic mechanical safety.
The domestic AI story is therefore not simply adoption: users want agents to act, but companies must earn trust around permissions, data, reliability, and prices.
Chinese coverage places household convenience, hardware scarcity, automotive quality, and AI governance in one consumer narrative, a useful contrast to model-centric Western coverage.

Open at 36Kr Ask Claude ChatGPT

China unveils an AI robot for wheat breeding

# Signal SCMP 5h

China has shown a robot prototype designed to help breeders select wheat strains from large hybrid populations, linking AI, agriculture, and food-security policy.

The robot is intended to reduce the manual burden of identifying useful traits among tens of thousands of hybrid offspring.
Its public debut at a wheat-breeding seminar places the project inside China's broader state-backed push to combine agricultural science with automation.
The source does not specify the robot's sensors, model, field autonomy, or measured improvement over expert breeders, so this is an early prototype rather than a demonstrated production system.
The strategic value would be highest if the system can close the loop from phenotype measurement to breeding decisions across many growing environments.

Open at SCMP Ask Claude ChatGPT

Huawei and Qualcomm sign a broad patent deal as Chinese AI finance accelerates

# Analysis TMTPost 4h

A Chinese weekly roundup pairs Huawei's cross-licensing deal and overseas AI distribution with fundraising, safety, and enterprise-agent developments across the market.

Huawei and Qualcomm reportedly agreed to cross-license portfolios covering 5G, computing, AI, and networking, while Qualcomm will acquire some Huawei US patents.
The arrangement signals managed coexistence in a strategically sensitive technology relationship: licensing and patent sales can continue even as supply-chain restrictions remain.
The roundup says Moonshot AI is nearing a pre-IPO financing at a reported $50 billion valuation and that Zhipu is using overseas cloud revenue sharing to distribute its model.
Those financing and distribution claims are reported rather than independently verified; the more durable signal is that Chinese AI firms are looking for monetization outside a single domestic platform.
Chinese coverage places the Huawei deal and AI commercialization inside one national technology narrative, rather than treating patents, models, and cloud channels as separate stories.

Open at TMTPost Ask Claude ChatGPT

Chinese appliance makers add AI labels as demand stays weak

# Culture TMTPost 6h

Chinese appliance coverage says AI features are spreading while overall replacement demand remains soft, making the label as much a margin-defense story as a technology story.

The article describes weak refrigerators, washing-machine, and air-conditioner demand, with trade-in subsidies pulling purchases forward rather than creating a broad recovery.
Useful applications include load-sensing washers and occupancy-aware air conditioners, while voice layers on simple appliances are presented as differentiation theater.
Retailers are staging interactive AI displays in stores because the experience is difficult to communicate online and helps bring cautious consumers into physical locations.
The Chinese market frame is important: when volume stalls and specifications converge, manufacturers compete through a new story, and AI supplies that story across categories.
The result is a split market in which AI can remove real friction in complex appliances but mostly adds interface and failure modes in simple ones.
For product teams, the lesson is to test whether intelligence reduces a recurring decision or merely creates another screen and cloud dependency.

Open at TMTPost Ask Claude ChatGPT

Chinese auto supplier Zhongding raises up to $270 million for robots and cooling

# Signal TMTPost 1h

Zhongding plans to put about $103 million of a convertible-bond raise into robot joints, betting that manufacturing know-how can bridge automotive components and emerging AI infrastructure.

The proposed 1.92 billion yuan raise is roughly $270 million; 732 million yuan, or about $103 million, would fund robot-joint production.
Planned capacity includes 150,000 harmonic reducers, 150,000 sensors, and 280,000 robot joints, while the company says current robot products remain small-batch.
Another 582 million yuan, roughly $82 million, targets thermal-management systems for electric vehicles, storage, and data centers.
The strategic logic is vertical adjacency: precision machining, materials, and quality control from auto supply chains can be reused in robotics and liquid cooling.
The source reports first-half revenue of 10.091 billion yuan, roughly $1.42 billion, with net profit down 16.74%, which explains the search for a second growth curve.
The investment case depends on customer certification and utilization; the filing estimates annual depreciation and amortization could rise by as much as 122 million yuan, roughly $17 million, after ramp-up.

Open at TMTPost Ask Claude ChatGPT

BYD rolls its 150,000th new-energy commercial vehicle off the line

# Signal Leiphone 10h

BYD's 150,000th commercial EV is a 600 kW electric tractor, signaling a push from urban buses and light trucks into China's harder long-haul freight market.

The Q3 tractor uses BYD's commercial-vehicle battery, dual-gun fast charging, and a peak motor output of 600 kW, targeting heavy-duty transport rather than short urban routes.
BYD says its commercial vehicles now reach more than 70 countries and that it is adding charging, fleet software, and lifecycle maintenance to vehicle sales.
The milestone is strategically more important than the round number because it lands on a long-haul truck, where range, payload, charging time, and uptime make electrification harder.
The company is using vertical integration to sell a complete zero-carbon logistics system, not simply a battery vehicle.
The source is corporate material and provides no delivered fleet economics, so the commercial breakthrough remains an industrial ambition rather than a proved replacement for diesel on every route.

Open at Leiphone Ask Claude ChatGPT

Open AI systems are still hard to reuse reliably

# Analysis InfoQ China 0m

A Chinese analysis argues that open weights and code are not enough: reproducible AI requires compatible models, data, tools, environments, interfaces, and evaluation traces.

The article frames reuse as a systems problem: changing a model, tool interface, prompt, or runtime can shift an agent's behavior even when the surrounding code is unchanged.
For enterprise agents, state management, recovery, permissions, and multi-agent coordination matter as much as model quality, making a runnable demo a weak measure of portability.
It cites a 2026 open-source AI competition as evidence that Chinese developers are increasingly judged on deployability and closed-loop execution, not only benchmark scores.
The Chinese framing is notably implementation-focused: open source is treated as an infrastructure and ecosystem question, aligned with domestic efforts to standardize agent tools and interfaces.
The strategic gap is measurement. Stars, downloads, and single-run benchmarks do not show whether an independent team can deploy, audit, and maintain the system.

Open at InfoQ China Ask Claude ChatGPT

China uses dancing robots to court African business audiences

# Culture SCMP 6h

Chinese embassies used dancing robots and robot dogs at National Day events in Africa, presenting robotics as a compact symbol of industrial upgrading and partnership.

The events in Angola and Congo were diplomatic showcases, not reported commercial deployments, but officials explicitly linked the demonstrations to China's move toward innovation and higher-quality growth.
The robot performances turn an abstract industrial policy into an accessible cultural display for local officials, companies, and the public.
This is soft power with a supply-chain objective: China wants African partners to associate its technology exports with visible modernity, not only construction and commodity trade.
The demonstrations reveal little about technical capability, but they show how robotics is being packaged as a diplomatic language alongside trade and infrastructure.

Open at SCMP Ask Claude ChatGPT

Xi and Modi cannot close the India-China divide alone

# Analysis SCMP 12h

A Chinese analysis sees useful signals in Xi-Modi talks on open-weight AI but argues that unresolved structural tensions still outweigh diplomatic warmth.

The article treats cooperation on open-weight AI as a practical area where India and China could capture mutual benefit without resolving every geopolitical dispute.
It contrasts that signal with persistent concerns about an asymmetric relationship, showing how economic and strategic narratives compete in Chinese coverage.
The broader lesson is that leader-level engagement can reopen channels but cannot by itself repair border, trade, security, and trust problems.
For technology companies, open models may become a narrow bridge even while broader India-China commercial ties remain politically constrained.

Open at SCMP Ask Claude ChatGPT

LaserCyber raises millions for AI-assisted desktop laser welding

# Signal 36Kr 11h

Shenzhen startup LaserCyber raised several tens of millions of yuan, roughly $4-14 million, to turn industrial laser welding into a guided tool for overseas workshops and makers.

Its L1 machines cover 600W, 900W, and 1,200W and combine welding, cutting, and cleaning in one platform, including processing of reflective metals such as copper and aluminum.
The company is embedding process data and a smaller model into the machine to recommend parameters for material, thickness, power, temperature, and cooling conditions.
That architecture lowers the skill barrier without removing expert controls: novice users get guided setup while experienced operators can adjust a larger parameter space.
The product is a Class 4 fiber-laser system, so market expansion depends on safety interlocks, thermal design, and regulatory approval in Europe and the United States as much as on software quality.
The company says its first product has raised more than $1 million on Kickstarter, but the source does not provide production volume or field-failure data.
The strategic opportunity is industrial democratization; the constraint is that AI can recommend a weld, but it cannot make optical and mechanical hazards disappear.

Open at 36Kr Ask Claude ChatGPT

A new Taipei-Hong Kong subsea route is added for resilience

# Analysis SCMP 13h

A network operator added another Taipei-Hong Kong route after repeated cable disruptions, linking Taiwan's connectivity resilience with growing demand from AI traffic.

The new route brings the operator's Taiwan service to six connections: four between Taipei and Hong Kong and two between Taipei and Tokyo.
The practical strategy is path diversity rather than a claim that submarine cables are no longer vulnerable; multiple routes reduce the impact of a single break.
Hong Kong's role as the diversification hub reflects its continuing importance as a regional connectivity and data-center node despite wider geopolitical tension.
The AI angle is demand-side: inference and data traffic make connectivity redundancy a capacity and continuity issue, not only a telecom concern.
The source does not establish whether disruptions were accidental, environmental, or deliberate, so the security implications should not be overstated.

Open at SCMP Ask Claude ChatGPT

Two Qing-era ceramics damaged at the British Museum

# Culture SCMP 3h

Chinese ceramics were damaged during display at the British Museum, reviving questions about museum security and the stewardship of fragile Chinese heritage abroad.

The museum says a bowl fell from a shelf onto a small box, damaging both objects; its conservation team is restoring them.
The incident matters in China because overseas display of Chinese antiquities is often judged through a sovereignty and custodianship lens, not only a museum-safety lens.
The source does not establish whether the failure involved shelving, vibration, visitor access, or staff procedure, so broader security conclusions remain premature.
The episode may intensify pressure on museums to document display conditions and justify how they protect contested or culturally significant collections.

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	1/20
Ginger River Review	200	0/20
Sinocism	200	0/20
QbitAI	200	1/10
36Kr	200	5/30
Solidot	200	8/20
Leiphone	200	20/20
GeekPark	200	3/30
IT Home	200	12/60	+44 over cap
SCMP	200	10/50	+6 over cap
Initium	200	2/15
Jiemian	200	8/30	+22 over cap
China Digital Times	200	0/15
The Wire China	200	1/50
Techmeme	200	3/15
Hacker News	200	0/20
V2EX	200	8/50	+24 over cap
Zhidx	200	4/20
InfoQ China	200	8/20	+12 over cap
New Wisdom	200	8/15	+2 over cap
TMTPost	200	8/18	+10 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