China Desk — 22 Aug 2026, 04:42 UTC

Chinese chip startup lights up a four-layer 3D DRAM compute chip

# Leiphone 22h

Qianhe Yibang says its four-layer 3D DRAM compute-in-memory chip has returned from fabrication and powered on, moving a difficult architecture from research validation toward engineering deployment.

The chip uses a logic base with four vertically stacked DRAM layers, a 4-plus-1 structure intended to reduce data movement across the memory wall.
The company claims order-of-magnitude gains in bandwidth, latency, power, throughput and per-operation cost, but the supplied report provides no independent silicon measurements.
Vertical stacking makes thermal control, alignment, interconnect uniformity and yield the central engineering challenges; increasing layers raises all of them simultaneously.
The design targets high-concurrency workloads such as cloud gaming, where memory bandwidth and data movement can dominate accelerator economics.
Qianhe Yibang says it controls architecture, front- and back-end design, advanced packaging and manufacturing with domestic partners, aligning the product with China’s push for controllable semiconductor supply chains.
The startup recently completed a B round above 2 billion yuan (roughly $280 million), providing a substantial demand and financing base for commercialization.
The decisive test is whether the demonstrated chip can achieve stable yields, software support and competitive cost in volume rather than merely light up in the lab.

Open at Leiphone Ask Claude ChatGPT

Hong Kong court convicts Alliance leaders under national-security law

# Initium 21h

Chinese-language legal coverage reports convictions of Alliance leaders and the organization itself, with the court treating key individuals as its directing mind under Hong Kong’s national-security framework.

The court convicted former chair Lee Cheuk-yan and former vice-chair Chow Hang-tung, while another former vice-chair had pleaded guilty earlier.
The judgment applied an attribution doctrine so the organization’s alleged intent could be inferred through its senior leaders.
The charged offense carries up to ten years for serious cases and up to five years for lesser cases, with sentencing still pending in the supplied account.
The defense argued that the defendants had not used illegal means and lacked hostility toward the Communist Party, but the court rejected that account.
The verdict shows how Hong Kong’s national-security enforcement can reach an organization’s historical purpose and leadership intent, not only a discrete act of violence.
The case is a major marker in the narrowing space for public commemoration and civil-society organization after the 2020 security-law turn.

Hong Kong families mark 1,800 days around the Alliance case Initium

Open at Initium Ask Claude ChatGPT

Xi uses Jiang Zemin centenary to signal Party continuity

# SCMP 16h

The centenary memorial for Jiang Zemin, attended and addressed by Xi Jinping, recasts Jiang’s reform-era legacy inside the Party’s continuous modernization narrative.

The event was more than a commemoration: it placed Jiang’s period of leadership and its reform legacy within the current Party story.
By presiding over the memorial, Xi can acknowledge economic opening without conceding an alternative political direction or leadership tradition.
The framing suggests that China’s official history is being curated to connect different leadership eras rather than present them as competing models.
This is especially significant because Jiang’s period is associated with market reform, global integration and a different style of elite politics.
The memorial offers a controlled signal to officials and the public that historical continuity, not ideological rupture, is the preferred interpretation of the Party’s evolution.
Its meaning depends on what the event omits: the source does not describe any policy concession or revival of Jiang-era political pluralism.

Open at SCMP Ask Claude ChatGPT

Reusable Chinese rockets could strengthen the PLA satellite kill chain

# SCMP 15h

Recent booster recoveries move China closer to reusable launch, which analysts say could make the PLA’s satellite-dependent targeting chain more resilient in conflict.

The reported recoveries include a stainless-steel booster returned after an orbital flight, a first for China according to the source.
Reusability can shorten launch turnaround and lower the cost of replacing or augmenting surveillance, communications and positioning satellites.
That matters militarily because a kill chain linking detection, tracking and engagement depends on space assets that an adversary could try to disrupt.
The commercial launch sector therefore has dual-use significance even when a flight is not a military mission.
The key uncertainty is operational tempo: one or two recoveries do not prove aircraft-like reuse, rapid refurbishment or wartime surge capacity.
The strategic effect becomes larger if China can combine reusable boosters with distributed satellite constellations and resilient ground infrastructure.
The source relies on analyst assessment, so the military implications remain contingent rather than demonstrated.

Open at SCMP Ask Claude ChatGPT

China’s trade surplus with most EU states widens before October talks

# SCMP 14h

New customs calculations show China’s surplus expanded with 24 of 27 EU members in July, sharpening Brussels’ leverage problem before an October negotiating deadline.

The source reports especially large year-on-year swings with Sweden and Malta, though it does not provide a full explanation for either change.
The figures suggest that European efforts to rebalance trade are not yet changing the aggregate direction of goods flows.
For Beijing, persistent surpluses can support industrial scale but also intensify political demands for market access, procurement reciprocity and import controls.
For Brussels, the timing matters: negotiations are proceeding while the underlying trade data continues to move against its stated objectives.
The calculation is based on Chinese customs data, so definitions and the treatment of re-exports should be checked before drawing sector-level conclusions.
The broader signal is that trade friction is becoming a structural relationship problem, not a temporary negotiation gap.

Open at SCMP Ask Claude ChatGPT

Wang Yi’s South Korea visit leaves Seoul’s peace initiative unsupported

# SCMP 2h

Chinese coverage presents Wang Yi’s Seoul visit through China’s North Korea alignment, while the regional account stresses Seoul’s disappointment that Beijing would not back multiparty peace talks.

The visit followed a recent China-North Korea leadership exchange and coincided with uncertainty over US military exercises with South Korea.
The key diplomatic gap is that Seoul sought a broader peace-talk format, while China did not endorse that proposal in the supplied account.
For Beijing, maintaining strategic space with Pyongyang appears to outweigh helping Seoul build a multilateral process around the peninsula.
The episode shows how China can engage South Korea extensively without allowing that engagement to weaken its North Korea channel.
The outcome matters because Washington’s shifting posture creates room for Seoul to seek more autonomy, but also makes Chinese and North Korean coordination more consequential.

Open at SCMP Ask Claude ChatGPT

Robot model learns new tasks from seconds of demonstration

# AI QbitAI 21h

Generalist’s GEN-1.5 model reportedly learns novel robot actions from 3-to-12-second demonstrations without gradient updates, bringing in-context learning into physical manipulation.

The model can combine separate demonstrations and transfer simulated demonstrations to a real robot, treating motion examples as a physical prompt.
In tests across ten tasks, one-shot success averaged 59%; adding five minutes of data and ten gradient updates raised it to 83%.
The key architectural claim is that the ability emerged from large-scale physical-data pretraining rather than a specially designed meta-learning objective.
This is analogous to language-model in-context learning, but the action context must also account for contact dynamics, embodiment and recovery from errors.
The current tasks are short and atomic, and the source acknowledges that fine-tuned models remain more stable.
A 59% result is not production-ready, but achieving it with zero updates is a meaningful signal that physical pretraining can create new adaptation behavior.
The next test is whether the same mechanism survives long-horizon tasks, different bodies and safety-critical contact.

Open at QbitAI Ask Claude ChatGPT

Qianxun shows why embodied AI progress depends on data loops

# AI QbitAI 23h

Qianxun repeated a living-room cleanup demo at China’s robotics conference to show month-on-month improvement in long-horizon tasks, including an 85% generalized trash-picking success rate.

The task requires perception, navigation, object state tracking, grasping, door interaction and recovery, so a small error can propagate through many later actions.
Qianxun attributes the progress to a pipeline combining web video, first-person data, teleoperation, real-robot logs, labeling, quality checks and evaluation.
Its seventh-generation collection hardware reportedly cuts collection cost to one-tenth of traditional teleoperation and raises usable data from 30% to 95%; these are company claims.
The company’s idea of training on dirty data reflects a practical view of physical deployment: lighting, clutter, occlusion and human traffic are part of the target distribution, not noise to remove.
The technical architecture uses an agent for task memory and decomposition, navigation for positioning and a base model for action generation, with post-task reports feeding the next training cycle.
Repeating the same demo is valuable only if it measures system evolution rather than scripted presentation; the source itself warns that show-floor success does not prove general reliability.
The decisive commercial metric will be intervention rate and cost over many real tasks, not a single 15-minute demonstration.

Open at QbitAI Ask Claude ChatGPT

Chinese robot completes 15 minutes of unscripted household work

# AI QbitAI 1d

Moqi demonstrated a wheeled robot completing a 15-minute household sequence without human instruction, using an agent-native policy model to maintain goals, memory and recovery inside the action loop.

The task included clearing a table, checking and replenishing a refrigerator, moving wet laundry to a dryer and folding it afterward.
MoRA differs from a conventional planner-plus-action-policy stack by putting goal conditioning, execution memory and progress awareness inside the lower-level policy.
The proposed system separates a project-manager-like reasoning layer from a worker-like action layer that can continue operating and call for help only at its boundary.
This architecture addresses a practical failure mode of long-horizon robotics: repeated pauses while a high-level model replans every short action.
The source says the robot was trained to accept structured text-and-image goals and maintain multiple levels of progress memory, but provides no intervention or repeatability statistics.
The demo is valuable because home environments combine navigation, manipulation and changing object states, yet one exhibition run cannot establish household reliability.
The commercial bet is that fewer high-level interventions can make a general robot cheaper to deploy than a collection of narrowly scripted machines.

Open at QbitAI Ask Claude ChatGPT

China’s Z20 brings L4 autonomy to four-ton industrial trucks

# AI QbitAI 16h

Ninebot-style autonomy specialist Jishi and Yutong launched a 4.2-ton driverless truck for factories and logistics parks, shifting Chinese autonomous driving from last-mile delivery toward industrial throughput.

The Z20 carries 4.2 tons, offers 19.32 cubic meters of cargo space and can handle six to eight standard pallets, according to the company.
Jishi supplies the L4 perception, planning, control and fleet-management stack, while Yutong contributes the chassis, manufacturing and quality systems.
Battery options include 43.47-kWh and 77.28-kWh packs; the long-range version is claimed to reach 320 km empty and recharge from 30% to 80% in about 30 minutes.
The product is designed around the whole workflow, including box bodies, hydraulic liftgates and loading constraints, rather than simply adding autonomy to an existing truck.
Closed industrial sites offer fixed routes, frequent trips and measurable labor savings, making them a more plausible early market than unrestricted urban roads.
The company’s reported 25,000-vehicle delivery base and 170 million kilometers of L4 operations are self-reported, not independently audited.

Open at QbitAI Ask Claude ChatGPT

Chinese startup launches an AI workspace for end-to-end science

# AI QbitAI 22h

Deep Research’s Bohr Science Space puts literature review, experiment planning, computation, analysis and paper checking into one desktop environment for Chinese research teams.

The workspace connects local and cloud data, models, software and compute resources, while discipline-specific agents target biology, medicine, drug discovery and materials science.
Its technical proposition is workflow continuity: an agent can retrieve evidence, execute code, inspect results and preserve context for the next scientific decision.
For biology, the system reportedly links papers, public databases and code to reproduce analyses rather than merely summarize them.
The product addresses a real bottleneck in research automation: provenance, environment setup and result validation are harder than generating a plausible paragraph.
The source describes a public beta and extensive capabilities but provides no independent benchmark of reproducibility, error rates or laboratory integration.

Open at QbitAI Ask Claude ChatGPT

RayNeo iO targets all-day AI glasses at about $281

# AI QbitAI 19h

Chinese AR brand RayNeo launched iO from 1,996 yuan (about $281), combining a 34-gram frame, two-day battery claim and proactive memory features to make AI glasses a daily device.

The glasses use a three-layer waveguide, a tiny optical engine and a 33-inch-equivalent display, with claimed 93% transparency in the optical-combiner area.
RayNeo’s software stack supports multiple Chinese open models and a memory engine that records contextual information from sound, location and daily activity.
The product also includes a physical crown, head gestures and translation across 55 languages and 109 accents, showing an attempt to reduce phone dependence through low-friction control.
The privacy trade-off is central: an always-on device that summarizes conversations and supplies unsolicited prompts needs clear retention, consent and deletion controls, none of which the supplied text details.
The significance is positioning: a Chinese AR vendor is treating model choice and continuous context as product infrastructure rather than attaching a chatbot to conventional glasses.

Open at QbitAI Ask Claude ChatGPT

Alibaba’s AI strategy is becoming a full model portfolio

# AI GeekPark 18h

Alibaba’s latest results present AI as a systems business: a model portfolio across text, image, audio, video and music tied to cloud growth, open-source adoption and infrastructure scale.

Also: Leiphone, 36Kr

Alibaba Cloud external commercial revenue rose 45% to 48.4 billion yuan (roughly $6.8 billion), while adjusted EBITA rose 133% to 5.6 billion yuan (about $790 million).
The portfolio spans Qwen language models, image and audio systems, Wan video models and a new music model, shifting competition from one model to coordinated production capacity.
Alibaba says Qwen models have exceeded 3 billion downloads and generated more than 300,000 derivatives, a distribution advantage that matters more than a single benchmark win.
The reported 2.4-trillion-parameter Qwen model and 27-billion-parameter open model illustrate a portfolio strategy spanning frontier scale and deployable efficiency.
The strategic claim is that training infrastructure, data, model release cadence and ecosystem adoption form a factory-like capability that rivals cannot easily copy one product at a time.
The figures come from company and outlet reporting, so benchmark rankings and derivative counts should be independently checked before treating them as durable lead indicators.

Open at GeekPark Ask Claude ChatGPT

Study says China could meet only 60% of wartime drone demand

# Signal SCMP 18h

An AI-assisted supply-chain study estimates China’s industrial system could satisfy only 60% of military drone demand in a sudden conflict, revealing a wartime scaling constraint beneath manufacturing capacity.

The model mapped drone assemblers to subsystem and component suppliers while accounting for inventories, backlogs, budgets and civilian demand.
The 60% estimate implies that nominal production capacity is not the same as surge capacity when a conflict disrupts inputs and redirects factories.
The result highlights bottlenecks that may sit below final assembly, including components, inventories and supplier concentration.
Because the estimate comes from a model rather than observed wartime output, its value is scenario analysis, not a forecast of actual combat supply.
The study is important precisely because it challenges the simplistic view that China’s large manufacturing base automatically guarantees unlimited drone replenishment.

Open at SCMP Ask Claude ChatGPT

China and Indonesia deepen military ties after joint naval drill

# Policy SCMP 16h

Chinese coverage treats the joint drill east of Taiwan and a ministerial meeting as a foundation for wider cooperation, while regional observers debate whether it lowers the political barrier for other Southeast Asian states.

Also: Jiemian

The exercise was followed by talks between the countries’ foreign and defense ministers covering military cooperation and food security.
Indonesia’s participation matters because it has traditionally sought room to maneuver among major powers rather than formal alignment with China.
For Beijing, the activity demonstrates that security cooperation can expand beyond treaty allies and into practical exercises with a large Southeast Asian partner.
The source says observers believe other regional governments may see the drill as a precedent, though it does not show that any have decided to follow.
The event’s significance lies less in its immediate combat value than in normalizing a broader China-Indonesia security relationship.

Indonesia’s president meets China’s foreign and defense ministers Jiemian

Open at SCMP Ask Claude ChatGPT

China and the US compete for Southeast Asia’s AI alignment

# Policy SCMP 13h

Chinese analysis frames Southeast Asia as a contest between Beijing’s AI cooperation network and Washington’s China-free supply chain, testing the region’s non-alignment.

The US initiative seeks to bind regional partners into an AI supply chain that excludes China, while Beijing is promoting its own international AI cooperation structure.
For Southeast Asian governments, the choice is not simply about model access: chips, cloud infrastructure, standards and security relationships are bundled together.
The pressure exposes a tension between economic diversification and strategic autonomy, since joining either ecosystem can create long-term technical dependencies.
The article’s China-focused view is useful because it treats AI infrastructure as foreign policy rather than as a purely commercial market.

Open at SCMP Ask Claude ChatGPT

Solar expansion linked to a 2.1% drop in bird diversity

# Signal Solidot 1d

A study using data from 2,344 Chinese counties links stronger solar-development policy to lower bird diversity, exposing a land-use trade-off inside China’s clean-energy buildout.

The researchers combined bird observations, solar policy, environmental conditions and socioeconomic data from 2014 to 2023.
A one-standard-deviation increase in policy intensity corresponded to a 2.10% decline in the bird-diversity index, with stronger effects in wealthier and non-desert areas.
The proposed mechanism is rapid conversion of farmland and grassland into development zones, reducing vegetation diversity and fragmenting habitat.
The result complicates a common China energy narrative in which renewable expansion is treated as an unqualified environmental gain.
The study is observational and does not prove that every solar project causes the same effect; siting, construction practices and restoration could change the outcome.
For infrastructure planners, the practical implication is that biodiversity constraints need to be designed into renewable deployment rather than added after capacity targets are set.

Open at Solidot Ask Claude ChatGPT

Europe’s sovereign-AI push quietly depends on China’s Z.ai

# Policy SCMP 18h

Mistral’s sovereign-AI infrastructure pitch reportedly includes China’s Z.ai models, exposing a contradiction between European technology autonomy and dependence on Chinese model supply.

The issue is not simply model quality: offering an external Chinese model inside European infrastructure can weaken claims of control over the full AI stack.
For Mistral, a multi-model platform may be commercially rational because customers want capability and choice, even when governments want provenance and strategic independence.
The episode shows how sovereignty is layered: local data centers and European hosting do not necessarily mean local model ownership.
The source frames the arrangement as a strategic dilemma for the European Union, where decoupling rhetoric collides with the practical scarcity of competitive models and infrastructure.
The unresolved question is whether European customers will accept Chinese model components when the service is hosted and governed by a European company.

Open at SCMP Ask Claude ChatGPT

TikTok and ByteDance settle US children’s privacy case for $400 million

# Policy SCMP 7h

Chinese coverage stresses the settlement’s significance while TikTok’s US politics remain tense; the US account focuses on resolving a Justice Department privacy case for $400 million.

Also: Techmeme

TikTok and ByteDance agreed to settle allegations that the platform violated US children’s online privacy laws, according to the supplied report.
The settlement closes a major federal case but does not remove broader political scrutiny of TikTok’s US operations.
The China angle is structural: ByteDance’s parentage remains central to how US regulators treat the platform, even when the immediate allegation concerns child privacy.

US DOJ and TikTok reach $400 million children’s privacy settlement Techmeme

Open at SCMP Ask Claude ChatGPT

Baidu Cloud rebuilds its partner ecosystem around AI delivery

# Policy Leiphone 22h

Baidu Cloud says it is replacing loose project-by-project resale with a structured partner system built around industry expertise, AI products and delivery capacity.

The company now distinguishes product, opportunity, delivery and channel partners, allowing one partner to accumulate several capabilities over time.
The shift reflects a changed customer requirement: buyers want an AI solution tied to a business process, not merely compute or a model endpoint.
Baidu reports 79% quarterly cloud growth and says 10% of core partners grew faster than the platform, but one quarter is not proof of durable ecosystem health.
The three-year plan prioritizes restructuring first, validating repeatable scenarios second and scaling partner economics third.
The strategic issue is whether a cloud provider can give partners enough room to build differentiated products while still capturing platform value.
The source is an interview with a company executive, so the growth and partner figures should be treated as reported claims rather than independent market measurement.

Open at Leiphone Ask Claude ChatGPT

China prepares Chang’e 7 mission to search for lunar south-pole ice

# Signal Solidot 15h

China plans to send Chang’e 7 toward the lunar south pole, using a hopping vehicle to inspect permanently shadowed craters for water ice.

The mission is scheduled to launch on a Long March 5 from Wenchang, reach lunar orbit after about six days and prepare for a planned November landing.
Its architecture combines an orbiter, lander, rover and a hopping vehicle that can recharge in sunlight before entering shadowed craters.
The mission’s scientific value is direct sampling and observation: orbital evidence suggests polar ice, but its amount and distribution remain uncertain.
A successful hopping vehicle would also demonstrate mobility in terrain where conventional solar-powered rovers cannot operate continuously.
The project illustrates China’s preference for a sustained, hardware-rich lunar exploration sequence rather than a single payload mission.

Open at Solidot Ask Claude ChatGPT

Pop Mart shifts from hypergrowth to shareholder returns

# Culture 36Kr 17h

Pop Mart’s revenue still grew, but overseas sales fell 11% and inventory days rose to 201 as the Chinese collectible-toy company responds with buybacks, dividends and a less aggressive expansion stance.

First-half revenue was 17.17 billion yuan (roughly $2.4 billion), adjusted profit 5.16 billion yuan (about $730 million), and gross margin 69.7%.
Overseas revenue fell to 4.97 billion yuan (about $700 million), while inventory rose to 6.1 billion yuan (roughly $860 million) and turnover days increased from 123 to 201.
The overseas slowdown exposes the risk of building global demand around one viral character: the Monsters family remained largest, but its share fell as Starry Night rose rapidly.
Pop Mart says it will prioritize profit over growth and plans 2 billion to 5 billion yuan (roughly $280 million to $700 million) in repurchases.
Combined with about 3.2 billion yuan (roughly $450 million) in dividends, the payout signals management is using capital returns to stabilize investors during a growth reset.
The strategic test is whether Pop Mart can become a diversified IP platform rather than a company whose international economics depend on the lifecycle of one character.

Open at 36Kr Ask Claude ChatGPT

Bilibili relaunches an international app without identity checks

# Culture Solidot 21h

Bilibili is relaunching an international app and hiring community staff abroad, but its expansion will face the same moderation, data-security and political-friction problems as TikTok.

The new app reportedly allows registration without passports or national identity documents, lowering a barrier that limited earlier international access.
Hiring in Los Angeles, London, Mexico City, Sao Paulo, Istanbul and Tokyo suggests the company is building local community operations rather than relying only on translation.
Bilibili’s Chinese content base and outside creator partnerships give it a potential supply advantage, but global growth requires moderation norms that may differ sharply from its home market.
The platform’s challenge is not only acquiring viewers: it must explain governance, data handling and content enforcement to regulators and users accustomed to other platforms.
The move illustrates how Chinese internet companies try to export community products while separating global operations from domestic identity and compliance systems.

Open at Solidot Ask Claude ChatGPT

Chinese e-bike makers face backlash over subscriptions for smart features

# Culture Jiemian 14m

Owners of Chinese electric scooters say location, theft protection and remote control become paid services after a year, turning an upfront sale into a contested recurring-revenue model.

Reported annual fees range from 42 to 89 yuan (roughly $6 to $13), but consumers object more to unclear disclosure than to the absolute amount.
The dispute exposes a transition in China’s low-cost vehicle market from one-time hardware sales toward connected-service monetization.
Users argue that tracking and remote control were marketed as core product features rather than optional subscriptions, creating a gap between sales expectations and contract terms.
The source names Tailing, Niu, Ninebot and Yadea, suggesting the issue is broader than one company, though it does not establish a uniform industry policy.
The regulatory question is whether connected functions must be priced and disclosed like a durable vehicle feature or can be treated like a separate software service.

Open at Jiemian Ask Claude ChatGPT

Kuaishou bets more than $169 million on a new fantasy MMO

# Culture 36Kr 16h

Kuaishou’s costly Lord of Mysteries MMO launched into a cooling genre, using a major web-fiction IP and Unreal Engine 5 to test whether its self-developed games unit can break through.

The project cost more than 1.2 billion yuan (roughly $169 million), including marketing, and involved more than 600 staff.
Its design deliberately combines a faithful retelling of the novel with an original storyline set after the book’s ending, reducing the risk of alienating existing fans.
The game borrows familiar MMO progression and monetization patterns associated with successful Chinese titles, while using a Victorian and cosmic-horror aesthetic to differentiate the setting.
The commercial gamble is larger than launch rankings: Kuaishou must prove that it can turn an IP audience into durable retention and spending in a mature MMO market.
The source reports strong prelaunch reservations and an iOS free-chart peak, but those are not evidence of long-term revenue or player retention.

Open at 36Kr Ask Claude ChatGPT

WeChat voice transcription backlash meets DeepSeek and auto recalls

# Signal GeekPark 4h

Chinese tech coverage links a WeChat voice-transcription test, DeepSeek’s multimodal API and a 4.27-million-vehicle recall, reflecting a market where AI rollout and safety regulation arrive together.

Users criticized WeChat’s iOS-only gray test for making voice-to-text harder to discover or disable, showing how small interface changes can become mass online controversies.
DeepSeek’s new vision model accepts images for screenshot, chart and document understanding while preserving the text capability of its existing fast model.
A regulator-listed recall covers more than 4.27 million vehicles because emergency mechanical handles may be hard to see or operate after low-voltage failure.
The roundup also reports ByteDance restructuring its Seed model group into data, reinforcement learning, business post-training and consumer post-training units.
Because the source is a multi-story roundup, the individual claims are more useful as signals of Chinese product and regulatory priorities than as one unified event.

Open at GeekPark Ask Claude ChatGPT

A researcher warns that agents face a sim-to-real reliability gap

# Analysis Leiphone 1d

ASU researcher Hua Wei argues that language agents moving across languages and environments face the same observation, dynamics and reward gaps that break reinforcement-learning systems in real streets.

The framework divides the gap into observation, state, action, dynamics and reward differences, extending a robotics diagnosis to software agents.
Wei’s earlier traffic-control work in China is used as a concrete example: weather, human behavior and imperfect sensors can invalidate a policy that looked optimal in simulation.
The proposed remedy is domain randomization across prompts, actions and reward spaces; the source says a 3-billion-parameter model beat a 32-billion-parameter model under one cross-language test.
Uncertainty estimation and human-in-the-loop escalation are presented as safety mechanisms, especially when agents scale beyond what humans can supervise directly.
The central engineering lesson is that larger models do not automatically solve distribution shift; reliability depends on modeling what changes between training and deployment.
The reported result comes from a conference presentation and needs independent replication across tasks before it can support a general scaling claim.

Open at Leiphone Ask Claude ChatGPT

Chinese police dismantle cross-border stock-tip fraud network

# Policy Jiemian 8m

Chinese police say a Cambodia-based stock-tip scam caused one victim losses nearing $700,000, exposing a full pipeline from software and promotion to laundering.

The reported loss was nearly 5 million yuan (roughly $700,000), and 82 suspects were returned to China, including 25 fugitives.
The case illustrates how Chinese authorities frame cross-border fraud as an industrial chain rather than isolated deception, linking technical development, lead generation and payment laundering.
Cooperation with Cambodian police was central to the arrests, showing that enforcement increasingly depends on overseas operational partners.
The source presents this as a complete-chain crackdown, but gives no evidence about conviction rates, asset recovery or whether similar networks remain active.

Open at Jiemian Ask Claude ChatGPT

China leads the first international standard for solid-state EV batteries

# Policy IT Home 1h

An IEC proposal led by China has been approved as the first international standard project for automotive solid-state batteries, giving Chinese researchers influence over how the field is tested.

The proposed guide covers application, test items and conditions for secondary lithium-ion solid-state batteries used in electric vehicles.
Experts from France, South Korea and Japan are participating, so the move is standard-setting rather than a purely domestic specification.
The strategic value is control over comparability: common test methods can shape which claimed safety, range and performance gains become commercially credible.
The source does not show that the standard has been completed or that it resolves manufacturing and yield problems.

Open at IT Home Ask Claude ChatGPT

AliExpress reaches operating profitability through localization

# Signal Leiphone 1d

Chinese coverage says AliExpress reached operating profitability as local inventory and branded goods grew, presenting localization as the route from cross-border traffic platform to durable marketplace.

The platform’s managed overseas service reportedly covers more than 30 countries, with official European warehouses in Spain, France and Poland and another under construction in Germany.
Local orders exceeded half of volume in several core European markets during the cited promotion, while more than 20,000 overseas managed-service merchants joined in the first half.
The Brand Plus program reportedly pushed branded-buyer penetration above 30% and brand GMV toward 40%, shifting the proposition from cheap selection toward brand distribution.
The claimed profit is at the operating level, and the source gives no margin or cash-flow detail, so the durability of the improvement remains unclear.
The strategic question is whether Alibaba can make local fulfillment and brand trust strong enough to offset the regulatory and logistics complexity of global commerce.

AliExpress reports its first operating profit Leiphone

Open at Leiphone Ask Claude ChatGPT

Taiwan drills for mobile-network disruption affecting 17 million people

# Policy Initium 16h

Chinese-language coverage examines Taiwan’s exercise that reduced mobile speeds to 1% across 14 counties, treating communications resilience and legal authority as national-security questions.

The exercise ran alongside civil-defense drills and was designed to model communications degradation rather than a normal service outage.
Its scale makes network dependence visible to the public, including the need for alternate communication channels during a crisis.
The item focuses on Taiwan’s preparations and legal debate, not on a Chinese government action, so the China connection is strategic and indirect.
The source does not establish that the exercise simulated a specific attack or that the fallback systems would work under real interference.

Open at Initium Ask Claude ChatGPT

Electrochemist Liu Jiawei moves from Singapore to Hong Kong to start a lab

# Analysis SCMP 42m

A materials scientist moved from Singapore to Hong Kong to launch a City University lab, illustrating how Hong Kong can offer international research conditions while keeping closer mainland ties.

The source describes Liu’s move as both a family decision and a professional pivot into leading her own research group.
Her case reflects Hong Kong’s role as an intermediate research environment for scientists seeking access to mainland networks without leaving an international academic system.

Open at SCMP Ask Claude ChatGPT

East Buy says private-label products now drive most merchandise volume

# Culture Jiemian 34m

East Buy reported 36.3% revenue growth, with private-label goods accounting for 52.6% of merchandise volume as the Chinese livestream retailer builds a product business beyond traffic.

Revenue reached 20 billion yuan (roughly $2.8 billion), gross profit 2 billion yuan (about $280 million), and operating profit 660 million yuan (about $93 million).
Private-label products generated 5.4 billion yuan of GMV (roughly $760 million) out of 10.2 billion yuan (about $1.4 billion) overall.
The company launched 1,009 private-label products, indicating a shift from influencer-led distribution toward a more controlled assortment and margin structure.
Its new Hangzhou livestream studios show the operating model expanding by category rather than relying only on a single personality or channel.

Open at Jiemian Ask Claude ChatGPT

Lynk and Co recalls vehicles to replace defective lidar units

# Policy IT Home 1h

Lynk and Co is voluntarily recalling selected Chinese vehicles after supplier process variation damaged lidar power chips and could limit advanced driving functions.

The company says affected lidar units may trigger speed restrictions in assisted-driving modes, while ordinary driving remains available without those modes.
The remedy is free lidar replacement under the vehicle warranty, with the recall filed with China’s market regulator.
The failure mode matters because perception hardware is becoming a safety-critical dependency for highway and urban navigation systems.
Cloud diagnostics will help identify potentially affected vehicles, showing how connected-car infrastructure is becoming part of recall execution.
The source does not report crashes or injuries, and it attributes the defect to a supplier manufacturing fluctuation rather than the perception software itself.

Open at IT Home Ask Claude ChatGPT

ChinAI	403	0/0		HTTPError: 403 Client Error: Forbidden for url: https://chinai.substack.com/feed
ChinaTalk	200	1/20
Pekingnology	200	0/20
Ginger River Review	200	0/20
Sinocism	200	0/20
QbitAI	200	9/10
36Kr	200	9/30
Solidot	200	9/20
Leiphone	200	20/20
GeekPark	200	5/30
IT Home	200	12/60	+41 over cap
SCMP	200	10/50	+7 over cap
Initium	200	6/15
Jiemian	200	8/30	+20 over cap
China Digital Times	200	0/15
The Wire China	200	0/50
Techmeme	200	2/15
Hacker News	200	0/20
V2EX	200	8/50	+32 over cap

Two verdicts reached past the defendants — to the sons, and to the organisation itself

Life for Xu Jiayin, prison for his two sons

Hong Kong convicts a group, not only its leaders

A typhoon toll quadrupled six weeks late

Yangtze Memory wants $4.6bn at the top of the cycle

Clouds are now losing money selling Zhipu's models

Labubu cooled and the champions took over

Pekingnology audits the phrase "barbarian handler"

Threads we are pulling