China Desk — 21 Aug 2026

A robot learns new tasks from seconds of demonstration

# QbitAI

Generalist’s GEN-1.5 robot model reportedly learns unseen actions from 3 to 12 seconds of demonstration without gradient updates, an embodied analogue of in-context learning.

The model can combine separate demonstrations, transfer a simulated demonstration to a real robot and infer unshown transitions between subtasks.
That is architecturally different from ordinary fine-tuning: the recent physical demonstration becomes context for action rather than a new parameter update.
In tests across ten tasks, one-shot success averaged 59%; five minutes of additional data and ten gradient steps raised it to 83%.
The tasks remain short and relatively atomic, and the authors say in-context performance is less stable than fine-tuned policies.
The striking result is not the 59% alone but that the capability reportedly emerged during large-scale physical-data pretraining without a dedicated meta-learning objective.
If the approach scales to long-horizon and safety-critical work, robot deployment could become more like teaching by example than collecting a new dataset for every task.

Open at QbitAI Ask Claude ChatGPT

MoRA robot completes 15 minutes of household work without help

# QbitAI

Moqi’s MoRA architecture lets a household robot maintain goals, memory and progress inside the action model while completing a 15-minute multi-step chore sequence unattended.

The demonstration covers rubbish collection, refrigerator restocking, moving wet clothes to a dryer and folding them, a longer task than most public robot showcases.
MoRA makes System 1, the action policy, condition on structured text and image goals rather than only receiving short textual subtasks from a planner.
Its internal memory tracks immediate actions, subtask state and overall progress, allowing it to correct small deviations without repeatedly invoking the higher-level model.
System 2 remains responsible for intent, global planning and exceptions, but System 1 becomes an active worker instead of a reflexive motor controller.
This hybrid design addresses a common failure mode in embodied systems: frequent pauses caused by re-planning and loss of context between actions.
The source is company-led and the exhibition setting is controlled, so the open question is whether the architecture survives unfamiliar homes, safety constraints and long-tail failures.
If it does, the advantage is not a single impressive demo but a lower supervisory burden for household deployment.

Open at QbitAI Ask Claude ChatGPT

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

# Leiphone

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

The design stacks four DRAM layers over a logic layer, placing computation close to memory to attack the bandwidth and energy costs of moving data between separate units.
The company claims order-of-magnitude gains in bandwidth, latency, access power, throughput and per-operation cost, though no independent measurements are provided.
Adding layers increases alignment, yield, thermal and reliability problems, so a working sample is meaningful but not equivalent to volume production.
The startup says it controls the architecture, design, packaging and domestic manufacturing chain with local partners.
It recently raised more than 2 billion yuan, roughly $282 million, giving it a substantial runway to push from demonstration into products.
The initial target is highly parallel workloads such as cloud gaming, where memory bandwidth can dominate system performance.
The key follow-up is whether yield and thermal behavior hold across larger wafers and whether customers adopt the chip outside controlled demonstrations.

Open at Leiphone Ask Claude ChatGPT

China’s typhoon death toll rises to 159 after six-week gap

# SCMP

Chinese authorities raised the official death toll from Typhoon Maysak floods in Guangxi from 39 to 159, with 10 still missing and more than 1.65 million affected.

The revision came six weeks after the disaster, making the timing and process of official casualty reporting part of the story.
Authorities described the rainfall as historically rare, large-scale and persistent, pointing to disaster-management stress beyond a single storm event.
The affected population and revised toll show how infrastructure, local reporting and central statistics interact during a major Chinese disaster.
The source does not explain why the figures changed so sharply or whether an independent assessment is available.
For companies operating in southern China, the event is a reminder that climate and flood exposure can propagate through labor, logistics and industrial supply chains.
The next question is whether the post-disaster response produces changes in flood defenses, land use and reporting accountability.

Open at SCMP Ask Claude ChatGPT

Chinese cloud providers start selling rival models at a loss

# Leiphone

Chinese cloud sales teams reportedly discount third-party models such as GLM, Kimi and DeepSeek to as little as 30%, reviving the low-margin channel war of 2021.

Cloud companies are reselling models that customers specifically request because their own models are not always the preferred choice in coding and enterprise work.
When every provider sells the same model, discounts become the main differentiator; the report says 30% pricing can fall below cost.
The incentive problem is familiar: sales teams can book revenue by reselling an outside product even when it contributes little proprietary margin or capability.
That can inflate cloud revenue while weakening the economics of model development and creating conflict between product and sales organizations.
The difference from 2021 is that model companies may have more bargaining power and stronger customer pull, while cloud providers need their usage to hit AI targets.
The key question is whether providers eventually exclude low-margin resale from quotas, or accept losses to keep workloads and data inside their clouds.

Open at Leiphone Ask Claude ChatGPT

Xiaohongshu quietly open-sources a 280-billion-parameter model

# GeekPark

Xiaohongshu released a multimodal mixture-of-experts model with 280 billion total parameters, 16 billion active and a 512K context window, signaling that content platforms want their own AI substrate.

The dots3-note preview model handles text, images and speech and is released under Apache 2.0 through Hugging Face and GitHub.
Its design emphasizes search, agents and real-life decision tasks, matching Xiaohongshu’s data around travel, shopping, home projects and lifestyle choices.
The company also released benchmarks for multi-turn search and cross-stage life tasks, attempting to define evaluation around its strongest product context.
A same-day adaptation for Huawei Ascend hardware shows how quickly Chinese model ecosystems can connect open weights to domestic compute.
For a platform with hundreds of millions of users, self-hosting can reduce API exposure, control inference cost and keep data-to-product iteration inside the company.
The move also suggests model competition is spreading from AI specialists to any Chinese platform with high-frequency proprietary behavior data.
The model remains a preview, and the source reports gaps in complex programming and terminal tasks.

Open at GeekPark Ask Claude ChatGPT

China’s robotics expo shifts from model theater to closed loops

# 36Kr

Chinese reporting from WRC says buyers now care less about parameter counts and more about real work, with world models, fast and slow control loops and deployment data becoming the new industry vocabulary.

The expo reportedly drew more than 300 exhibitors, 3,000 robots and 49 state-owned enterprise groups carrying 12 real engineering scenarios, a stronger demand signal than a purely investor-led show.
Companies are moving beyond single VLA models toward world models, explicit or latent reasoning and closed loops that feed failures back into training.
A recurring engineering pattern is a fast local controller for repetitive motion plus a slower reasoning system for goals, planning and exceptions, reducing latency, cost and heat.
The source describes interfaces between models, navigation, hardware control and enterprise workflows as a central bottleneck, which explains why integration platforms matter as much as foundation models.
The most important commercial change is cultural: Chinese robotics companies are trying to show revenue-producing deployment, not merely intelligence in a controlled demonstration.
Exhibition claims about speed, data volume and success rates remain vendor-reported, so customer retention and intervention rates are the real next tests.

Open at 36Kr Ask Claude ChatGPT

China’s drone industry could meet only 60% of wartime demand

# SCMP

A supply-chain study estimates that even full Chinese mobilization would cover only 60% of military drone demand in a sudden conflict, exposing a gap between manufacturing scale and resilience.

The model maps drone assemblers to subsystem and component suppliers while accounting for inventories, production capacity, backlogs, budgets and civilian demand.
The finding challenges the assumption that China’s enormous commercial drone industry automatically converts into sufficient wartime supply.
A 40% shortfall could arise from bottlenecks in motors, sensors, batteries, flight controllers or other components rather than final assembly alone.
The result is a planning signal, not a battlefield forecast: the outcome depends on mobilization speed, stockpiles, imports and how demand is defined.
For the PLA, civilian production capacity is valuable only if critical nodes are diversified and can survive disruption or export restrictions.
The study’s use of an AI supply-chain model is itself notable, but its assumptions and validation matter as much as the headline percentage.

Open at SCMP Ask Claude ChatGPT

A Chinese robotics founder says VLA is not the endpoint

# Leiphone

At the World Robot Conference, Xingdong Era argued that world-action models should predict physical states as well as actions, moving embodied AI beyond imitation-based VLA systems.

VLA systems learn a powerful mapping from visual and language input to action, but their generalization is constrained by the human demonstrations in their training data.
A world model attempts to represent how objects and environments evolve, allowing the robot to reason about consequences that were not explicitly demonstrated.
The company’s proposed flywheel links a general model, a reusable robot body and real-world deployment data so that commercial use feeds model improvement.
The speech treats the robot body as part of the learning system, not merely an inference endpoint, because sensors, latency and mechanics shape the data distribution.
The argument aligns with a broader Chinese industry shift from isolated motion demos toward reusable models, deployment data and full-stack control.
World models remain a research direction; the report does not provide independent evidence that this approach outperforms VLA on real tasks.

Open at Leiphone Ask Claude ChatGPT

Jiang Zemin’s centenary becomes a signal of political continuity

# SCMP

Xi Jinping’s appearance at a Beijing memorial for Jiang Zemin uses party history to place Jiang’s reform-era legacy inside the current leadership’s story of modernization.

The memorial marked Jiang’s centenary but also offered a controlled interpretation of the period he led and its relationship to today’s Communist Party.
Chinese political commemorations routinely function as succession and policy signals, especially when they elevate a predecessor’s place in the official historical narrative.
The event suggests that continuity with reform-era modernization can be invoked without reopening the political limits of that era.
Its importance lies less in any announced policy than in who was present, who spoke and which parts of Jiang’s legacy were made usable for the current leadership.
The framing can reassure officials and investors that modernization remains a central party project while keeping historical interpretation under central control.

Open at SCMP Ask Claude ChatGPT

Reusable rockets could strengthen China’s military satellite resilience

# SCMP

Chinese coverage links recent booster recoveries to a possible PLA advantage: cheaper, more resilient access to the satellite networks used in targeting and battlefield coordination.

LandSpace’s recovery of a stainless-steel booster after an orbital flight is described as a first for China, alongside another recent recovery attempt.
Reusable launchers can improve launch cadence and reduce the cost of replacing satellites after attack, disruption or technical failure.
The strategic connection is the military kill chain, where sensing, communications and targeting depend on space infrastructure that an adversary could try to disable.
Commercial rocket companies therefore matter to national security even when their immediate business is civilian launch services.
The report cites analysts rather than military evidence; recovery is a milestone, not proof of routine reuse or wartime resilience.
The next meaningful indicators are rapid refurbishment, repeated orbital recoveries and integration with a larger domestic launch ecosystem.

Open at SCMP Ask Claude ChatGPT

China and Indonesia deepen defense ties after joint naval drill

# SCMP

A joint China-Indonesia naval exercise east of Taiwan was followed by a ministerial meeting, prompting debate over whether other Southeast Asian states may follow Jakarta’s lead.

The exercise may lower the political threshold for regional governments to conduct security activities with China without formally abandoning non-alignment.
Beijing’s message combines military cooperation with issues such as food security, presenting the relationship as a broad strategic partnership rather than a single defense pact.
The location gives the drill symbolic weight even though the report does not establish a new permanent military arrangement.
For Southeast Asian governments, the appeal is access and strategic room; the risk is being pulled into sharper US-China competition.
The next signal will be whether similar exercises become routine or remain a one-off demonstration of Indonesian autonomy.

Open at SCMP Ask Claude ChatGPT

Pop Mart shifts from explosive growth to shareholder returns

# 36Kr

Pop Mart’s revenue rose 23.8% to 17.17 billion yuan, roughly $2.4 billion, but overseas sales fell 11% as the company prepares to return much of last year’s profit to shareholders.

Adjusted profit increased only 9.5% to 5.16 billion yuan, roughly $727 million, while gross margin slipped to 69.7%, showing a sharp deceleration from the prior year’s expansion.
Overseas revenue dropped to 4.972 billion yuan, roughly $700 million, as supply constraints and weak brand depth outside Labubu exposed the cost of rapid international growth.
Inventory days rose from 123 to 201 and inventory reached 6.1 billion yuan, roughly $859 million, an operational warning behind the headline sales story.
Labubu’s family still generated 4.45 billion yuan, roughly $627 million, but its share fell as Star Person grew 580.6% to become the second major IP.
Management is prioritizing profit over aggressive expansion and plans 20 billion to 50 billion yuan, roughly $2.8 billion to $7.0 billion, in share buybacks.
The combined dividends and buybacks could consume roughly $940 million to $1.35 billion, a deliberate effort to stabilize investors while the company searches for a broader IP portfolio.

Open at 36Kr Ask Claude ChatGPT

Yangtze Memory starts a $46 billion-yuan Shanghai IPO

# IT Home

Yangtze Memory’s IPO application was accepted by the Shanghai exchange; it seeks 33 billion yuan, roughly $4.6 billion, to expand China’s leading 3D NAND maker.

The filing puts a major Chinese memory producer into the public-market financing pipeline after it reportedly returned to profit in 2024.
The company reported first-quarter 2026 revenue of 47.042 billion yuan, about $6.6 billion, and parent profit of 33.379 billion yuan, about $4.7 billion; those figures are unusually large and should be checked against the prospectus.
Yangtze Memory is controlled through Hubei state capital, making the listing both a financing event and a state-backed semiconductor milestone.
The company says it is among the global top three by NAND shipment capacity and operates two Wuhan fabs with combined monthly capacity of about 200,000 wafers.
The offering will reportedly fund a third production line, advanced-process research and working capital, so the strategic test is whether China can scale memory manufacturing under technology restrictions.
A successful listing would give domestic storage a deeper capital base; weak demand, yield or pricing would make the headline capacity less consequential.

Open at IT Home Ask Claude ChatGPT

Shareholder fight at Chinese foundry escalates into a seal dispute

# IT Home

A major shareholder says it took control of company seals after accusing a former chairman of unlawful control and 120 million yuan (roughly $17M) in related-party transfers, exposing governance risks at a Chinese foundry.

Hite High-Tech says it acted with other shareholders under court judgments, arbitration decisions and execution notices; the company previously alleged that more than 100 people forcibly entered and took its official seals.
The accusations include falsified board resolutions, disputed legal-representative changes and transfers through intermediary companies, but these claims remain the shareholder’s account rather than independently established facts.
The company reported first-half 2026 revenue below 60 million yuan (roughly $8.5M), suggesting that the governance conflict is occurring alongside serious operating deterioration.
In China, corporate seals can control contracts, banking and formal company actions, making physical possession unusually consequential in shareholder disputes.
The episode shows how ownership, court enforcement and party-company governance can collide in strategically sensitive semiconductor manufacturing.

Open at IT Home Ask Claude ChatGPT

Chinese team makes a micrometer-long single-atom metal wire

# IT Home

Researchers in Beijing produced stable carbon-encased copper atomic chains longer than one micrometer, extending the usable scale of one-dimensional metal structures by two orders of magnitude.

The chain is only one atom in diameter but contains more than 4,000 copper atoms, roughly analogous in aspect ratio to a four-meter household wire.
A high-pressure solid-state topological polymerization process creates a carbon protective sheath around the atomic chain, addressing the usual stability problem outside ultra-high vacuum.
The method reportedly extends beyond copper to cobalt, nickel and zinc, which matters more than the headline length if the synthesis is reproducible.
The work was published in Science and could support research into low-dimensional physics, nanoelectronics and flexible devices, although applications remain speculative.

Open at IT Home Ask Claude ChatGPT

Qianxun shows why repeating a robot demo can matter

# AI QbitAI

Qianxun repeated its living-room robot demo at Beijing’s robot conference, claiming an 85% success rate under varied object colors, types and positions after a month of iteration.

The long-horizon task combines navigation, perception, planning, grasping, placement and recovery, making it a better systems test than a single pick or gesture.
The company says its data pipeline combines web video, first-person data, teleoperation and real-robot records, with data quality rising from 30% to 95%.
Its stated training loop classifies why a task failed before feeding the case back into models, agents, navigation and hardware rather than simply collecting more demonstrations.
Qianxun promotes so-called dirty data because real homes contain clutter, changing light, occlusion and people crossing the robot’s path.
The report’s 85% number is company-reported and comes from exhibition conditions, so it does not establish household reliability.
The meaningful signal is iteration speed: if the same long task improves quickly through a closed loop, the company may have an operational advantage even without a new headline demo.

Open at QbitAI Ask Claude ChatGPT

China unveils prefabricated AI compute in shipping containers

# Signal QbitAI

Taichu Yuanki launched modular compute containers that it says can reach service within 24 hours, aiming to move Chinese AI capacity from central data centers into factories, mines and industrial parks.

The product line includes a 64-card air-cooled container, a 128-card liquid-cooled unit and a 256-card cluster rated at 307 petaflops of sparse FP16 compute.
The company claims more than 90% factory prefabrication and 70% faster delivery than conventional construction, with staged expansion and outdoor deployment.
A self-developed management platform is intended to schedule distributed nodes and expose training, inference and scientific-computing tools as one service layer.
The architecture reflects demand spreading from hyperscale cloud to smaller financial, energy, weather and industrial workloads that cannot wait for a new data center.
The reported 60% or higher utilization target and 24-hour commissioning are company claims, and power, cooling, networking and maintenance remain practical constraints.
If the model works, prefabricated compute could become a deployment pattern for Chinese industrial clusters constrained by real-estate, grid or permitting bottlenecks.

Open at QbitAI Ask Claude ChatGPT

Youdao reports $207M quarterly revenue as AI agents scale

# AI QbitAI

NetEase Youdao reported 1.47 billion yuan in second-quarter revenue, about $207 million, up 3.5 percent, while its AI agents expanded across education, translation, advertising and hardware.

The company says LobsterAI passed one million users and that it has remained operating-profitable for eight consecutive quarters, though the supplied account does not provide net income.
Learning services grew 20.9 percent to 800 million yuan (about $113 million), suggesting vertical AI is being monetized through an existing high-feedback business rather than a standalone chatbot.
AI subscription sales rose more than 20 percent to about 100 million yuan (roughly $14 million), while AI speaking-product sales more than doubled year over year.
Its advertising agents connect creator selection, regional data, content review and campaign generation, claiming to cut material-production cost by about 70 percent.
The strategic lesson is distribution: agents become commercially meaningful when embedded in products with existing users, workflows and feedback, not when sold as generic model access.
The figures are company-reported and mix revenue, usage and product claims; independent retention and gross-margin data are absent.

Open at QbitAI Ask Claude ChatGPT

Autonomous 4.2-ton truck targets Chinese industrial logistics

# AI QbitAI

Ninebot-style autonomous driving company Jiushi and Yutong launched a 4.2-ton driverless light truck for factory and warehouse transport rather than public-road delivery.

The Z20 carries 4.2 tons in a 19.32-cubic-meter cargo bay, moving the product beyond small last-mile delivery robots.
Ninebot-style developer Jiushi supplies the L4 driving stack and fleet management, while Yutong supplies the commercial-vehicle chassis and manufacturing system.
The target environments have fixed routes, repetitive tasks and measurable labor savings, making them more tractable than open-city autonomy.
Cargo loading, unloading and charging are treated as part of the product, showing that industrial autonomy is a workflow redesign rather than a sensor retrofit.
The company says it has delivered more than 25,000 autonomous vehicles and accumulated 170 million kilometers, but those figures are company claims.

Open at QbitAI Ask Claude ChatGPT

DeepModel launches a desktop environment for AI-assisted science

# AI QbitAI

Chinese AI company DeepModel launched a public beta that lets researchers ask questions while agents search literature, run analyses, call models and prepare papers.

The platform connects local and cloud data, computing and experimental resources rather than stopping at a conversational research assistant.
Domain agents cover biology, medicine, drug discovery and materials, with workflows that can retrieve sources, reproduce analyses and preserve intermediate results.
The design treats reproducibility and evidence checking as first-class functions, including citation review, chart validation and editable scientific output.
This is a more ambitious architecture than a chatbot, but the source is a product launch and gives no independent evidence of research productivity or error rates.
Its success depends on reliable tool execution, provenance and human review, not merely on the quality of the underlying language model.

Open at QbitAI Ask Claude ChatGPT

RayNeo launches lightweight iO AI glasses with two-day battery life

# AI QbitAI

Chinese AR company RayNeo launched iO glasses from 1,996 yuan, roughly $281, emphasizing ordinary eyewear, multimodel AI and an always-on memory assistant.

Also: IT Home

The glasses weigh 34 grams and use a three-layer waveguide, a small optical engine and a display claimed to reach 1,800 nits at the eye.
The product supports DeepSeek and Qwen models initially, with more models planned, making the device a model router as much as a single-assistant product.
Its memory engine processes audio and location context to generate summaries and proactive prompts, raising the more consequential question of continuous personal data capture.
A physical crown, head gestures, translation and planned links to payment, delivery and travel services show an attempt to make the glasses a daily interface.
The report is company-supplied and does not establish whether users will accept always-on sensing or whether battery claims hold in real use.

RayNeo shows a slim AI glasses design built around a tiny light engine IT Home

Open at QbitAI Ask Claude ChatGPT

UBTech's embodied model completes China's generative-AI filing

# Policy QbitAI

UBTech says its Walker embodied model passed the central cyberspace regulator's review, making it one of the first humanoid-robot models to complete China's generative-AI service filing.

The filing covers training-data production, safety labeling, red-team testing, safeguards, prompt engineering and agent design, extending model compliance into physical interaction.
UBTech tuned the system for emotional interaction while retaining perception and long-horizon planning, targeting home companionship, reception and education.
The company says its Thinker base model is open and has taken nine top benchmark results, but those performance claims are not independently verified here.
The regulatory milestone matters because humanoids in homes create a combined content-safety, privacy, physical-safety and anthropomorphism problem that ordinary chatbot filing does not fully capture.
Filing is permission to operate within a compliance framework, not proof that the robot is safe or commercially ready.

Open at QbitAI Ask Claude ChatGPT

Trump’s shortened Korea drills push Seoul toward defense autonomy

# Policy Initium

Chinese-language analysis says Trump’s abrupt reduction of US-South Korea exercises may accelerate Seoul’s push for wartime command transfer and more independent defense capacity.

The exercise ended four days early after Trump called it costly and an insult to North Korea, according to the source.
South Korean President Lee Jae-myung responded by linking operational-control transfer and a nuclear-powered submarine program to stronger self-defense.
The episode exposes an alliance-management problem: reducing exercises without advance coordination can undermine the very readiness assessments needed for Seoul to assume more responsibility.
The source frames autonomy and alliance as complementary rather than mutually exclusive, a politically useful position for Seoul amid uncertainty about US commitments.
This matters to China because changes in the US-South Korea command relationship affect deterrence, regional military planning and Beijing’s room for diplomatic maneuver.

Open at Initium Ask Claude ChatGPT

Hong Kong’s union verdict leaves supporters confronting a changed city

# Policy Initium

Chinese-language Hong Kong coverage centers on families and friends after the Alliance verdict, emphasizing the personal cost of nearly five years of prosecution and detention.

Three defendants, including former leaders Lee Cheuk-yan and Chow Hang-tung, were found guilty under the national-security sedition case; sentencing follows mitigation hearings.
The article treats the verdict as the closing of a long civic chapter, not simply a courtroom result, focusing on aging members, exile and the disappearance of public commemorations.
Family members described the ruling as unfair, while supporters framed their court submissions as a final effort to preserve a contested historical record.
The reporting also documents pressure on relatives and supporters who were arrested in separate incidents connected to remembrance activities.
The case shows how Hong Kong’s security regime reaches beyond formal defendants into networks of family, memory and public association.

Hong Kong court finds Alliance leaders guilty under security law Initium

Open at Initium Ask Claude ChatGPT

Alibaba assembles a full-stack model portfolio

# AI GeekPark

Chinese coverage of Alibaba’s earnings frames the company’s text, image, audio, video and music models as a coordinated industrial system rather than isolated benchmark wins.

Also: Leiphone

The reported portfolio spans Qwen language models, image and audio systems, Wan video generation and a new music model, shifting competition from one model to a production organization.
Alibaba says Qwen models have been downloaded more than 3 billion times and spawned more than 300,000 derivatives, evidence of ecosystem reach beyond its own applications.
The company’s cloud business reported 45% external commercial growth, linking open-model distribution to paid compute and enterprise services.
The strategic advantage, if the figures hold, is the ability to reuse training infrastructure, data pipelines and developer channels across modalities.
The report is highly promotional and relies heavily on company and benchmark claims; real enterprise retention remains the harder test.

Alibaba’s AI boom is making cloud infrastructure profitable Leiphone

Open at GeekPark Ask Claude ChatGPT

A Chinese researcher applies sim-to-real lessons to AI agents

# Analysis Leiphone

At IJCAI, Arizona State researcher Hua Wei argued that language agents face the same observation, dynamics and reward gaps that undermine simulated reinforcement-learning systems.

The talk connects earlier traffic-control work in complex Chinese roads with today’s agents that fail when moved across languages, tools or operating environments.
The proposed remedy is to randomize prompts, actions and rewards in training, borrowing domain-randomization methods from robotics.
The source says a 3-billion-parameter model using these perturbations outperformed a 32-billion-parameter model on some cross-language transfer tasks.
The lesson is architectural and operational: robustness may depend more on modeling environmental uncertainty than on simply increasing parameter count.
Hua also emphasizes uncertainty estimation and human-in-the-loop escalation so that agents ask for help without turning supervision into a new bottleneck.
The reported comparison needs independent reproduction, and sim-to-real analogies do not eliminate the need for domain-specific evaluation.

Open at Leiphone Ask Claude ChatGPT

KAI robot demonstrates autonomous full-match table tennis

# AI Leiphone

Chinese robotics firm KAI claims the first complete autonomous human-shaped-robot table-tennis match, using a world model and a full-stack data and deployment platform.

The SMASH 2.0 system links high-speed vision, trajectory prediction, whole-body control and action planning in millisecond-scale exchanges.
KAI says one policy can transfer across robot bodies, a meaningful test of whether embodied models are software assets rather than one-machine demonstrations.
The company’s KAI world model is trained on millions of videos and combines virtual-world generation, reconstruction and evaluation for sim-to-real transfer.
The hardware stack includes a 117-degree-of-freedom humanoid body, a 37-degree-of-freedom hand and a first-person data-capture headset.
These demonstrations are technically interesting, but the report provides no independent success-rate or production-deployment evidence.

Open at Leiphone Ask Claude ChatGPT

Chinese study links solar expansion to lower bird diversity

# Signal Solidot

A study using data from 2,344 Chinese counties finds that stronger solar-development policy correlates with a 2.10% decline in bird-diversity index per standard-deviation increase in policy intensity.

The dataset covers 2014 to 2023 and combines bird observations with solar policy, land use, vegetation and socioeconomic conditions.
The strongest effects appear in wealthier and non-desert areas, where farmland and grassland conversion may fragment habitats.
The result complicates a simple green-transition narrative: solar expansion can reduce emissions while imposing local ecological costs.
The study reports a correlation and attributes the mechanism primarily to rapid land conversion; it does not establish that every solar project has the same effect.
For Chinese planners, the policy implication is to treat siting, habitat protection and renewable deployment as one optimization problem rather than separate goals.

Open at Solidot Ask Claude ChatGPT

MiniMax turns its video model into an agentic design system

# AI GeekPark

MiniMax paired its H3 video model with a multi-agent design product, targeting end-to-end creative workflows rather than another standalone video generator.

H3 uses a three-stage design: a context and task-orchestration module, a 33-billion-parameter base generator and a 2K regeneration stage.
The context module can combine images, video and audio into structured instructions, addressing the prompt complexity and continuity problems that still limit video production.
MiniMax Design assigns work across agents for scripts, storyboards, video generation and assembly, shifting user effort from prompt engineering to intent specification.
The strategic bet is that generation quality will commoditize, while workflow integration and repeatable production become the durable product layer.
The source cites third-party rankings and creator examples but provides no independent cost or retention data.

Open at GeekPark Ask Claude ChatGPT

China fines Guotai Haitong arm $7.8 million over offshore funds

# Policy SCMP

Chinese regulators fined Guotai Haitong Asset Management 52.5 million yuan, about $7.8 million, for foreign-exchange violations in its offshore investment business.

The penalty includes 25.9 million yuan in fines and 26.7 million yuan in forfeited gains, according to the report.
It shows Beijing tightening control over outward capital flows through the qualified domestic institutional investor channel, even at a major brokerage group.
The enforcement matters for fund managers because offshore investment permissions are being treated as a regulated capital-allocation privilege, not simply a product-compliance issue.
The case does not establish a broad crackdown by itself, but it fits a wider pattern of scrutiny over foreign-exchange leakage and cross-border finance.

Open at SCMP Ask Claude ChatGPT

TikTok and ByteDance agree to a $400 million privacy settlement

# Policy SCMP

Chinese coverage foregrounds the US legal exposure around TikTok and ByteDance, while US coverage reduces the event to a $400 million children’s-privacy settlement.

Also: Techmeme

The settlement resolves Justice Department and Federal Trade Commission allegations that TikTok and ByteDance violated US children’s online-privacy law.
For TikTok, the payment closes a major federal case but does not remove the broader political scrutiny surrounding its US operations.
The China angle is structural: ByteDance remains the parent company, so ordinary privacy enforcement is entangled with the platform’s national-security and ownership debate.
The report does not specify the settlement’s operational requirements or admit wrongdoing, leaving the broader compliance impact unclear.
The difference in framing is useful: Chinese coverage treats the case as another pressure point on the Chinese-owned platform, while US coverage emphasizes consumer-law resolution.

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

Open at SCMP Ask Claude ChatGPT

A Chinese robotics founder pitches physical-world tokens

# Analysis Leiphone

Xinghaitu’s Gao Jiyang argues that embodied AI will eventually monetize completed physical tasks rather than robot hardware, treating machines as distribution for an action-based model economy.

The proposed progression is from selling robot hardware, to selling packaged solutions, to charging for effective physical actions or completed tasks.
Gao says hardware margins could fall from roughly 40% to 60% today, to about 20% for solution subscriptions and potentially negative in a token-led model.
The economic logic resembles cloud AI token pricing: the model provider captures recurring value while cheap devices place more inference endpoints in the world.
The company describes a unified autoregressive model for zero-shot generalization, general grasping and long-horizon work, supported by real-robot reinforcement learning.
The useful translation is that generalization equals training cost: every task a robot can perform without custom instruction reduces deployment expense.
The model is speculative; it requires stable success rates, measurable labor value and enough task volume to support recurring billing.

Open at Leiphone Ask Claude ChatGPT

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

# Signal Solidot

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

The mission is scheduled to launch on August 24 aboard a Long March 5 rocket, with landing planned for November after orbital preparation.
Its hopping vehicle is designed to move between sunlit areas and permanently shadowed regions, where water ice may survive but is difficult to access.
The mission architecture combines an orbiter, lander, rover and aerial hopping vehicle rather than relying on a single stationary surface platform.
Finding and characterizing accessible ice would inform future Chinese lunar infrastructure, but the mission will not by itself establish economically usable resources.

Open at Solidot Ask Claude ChatGPT

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