China Desk — 19 Aug 2026, 13:03 UTC

Chinese national lab puts humanoids to work running experiments

# QbitAI 6h

A national-level Chinese laboratory is using Monte2 humanoid robots for multistep liquid handling, cell work and toxicity testing, moving embodied AI from demonstrations into scientific production workflows.

The robots reportedly perform hundreds of operations across multiple instruments, including reagent transfer, mixing, pipetting and cell-toxicity experiments without continuous human operation.
The system is designed for long sequences and changing lab conditions rather than a fixed industrial pick-and-place routine, using vision, force and touch to adjust motion and grip.
The laboratory plans to expand to about 100 robots by the end of 2027 and coordinate them with AI, making this a test of fleet scheduling as well as manipulation.
The proposed model is a three-layer laboratory: a foundation model designs experiments, embodied robots execute them and the results feed back into analysis and planning.
The value proposition is higher throughput and consistency for repetitive biology work, while researchers spend more time on hypotheses and interpretation.
The hard test will be reproducibility, contamination control, recovery from failed steps and whether the system can generalize across protocols rather than only repeat validated recipes.

Open at QbitAI Ask Claude ChatGPT

Unitree's post-IPO plan is a self-improving robot engineering loop

# GeekPark 33m

On Unitree's listing day, founder Wang Xingxing described a system linking foundation models, code generation, simulation, real robots and evaluation so each deployment can improve the next.

Also: 36Kr

The proposed loop is search for papers and code, generate training programs, run simulation, deploy to hardware, score the result and feed the evidence back into another training cycle.
The important architecture is not a new robot model but an integrated development system in which AI participates in research, programming, simulation and testing.
Unitree's claimed advantage is physical distribution: its prospectus reports more than 30,000 cumulative quadruped shipments and over 5,500 humanoid shipments in 2025.
A large installed base could provide repeated real-world tests, but sold robots do not automatically become training data; consent, standardized telemetry and reliable data return remain unresolved.
The company also controls motors, joints, motion control and system integration, allowing hardware and policy changes to be debugged together rather than across supplier boundaries.
If Unitree can deploy thousands of machines and turn failures into reusable training data, its manufacturing scale could become an AI moat rather than merely a hardware advantage.
The proposal is still a roadmap, and its significance depends on real deployment volume, data rights and whether generated policies remain safe outside controlled demonstrations.

Open at GeekPark Ask Claude ChatGPT

A deleted post describes extreme security at Beijing's Tiananmen landmark

# China Digital Times 7h

Chinese censors removed a mundane account of identity checks and searches at a Beijing Tiananmen landmark, showing how broad enforcement can suppress even nonpolitical descriptions of security routines.

The deleted account reportedly described identity scans, pat-downs, bag and book checks and searches of clothing and shoes during a crowded summer visit.
The post did not criticize the measures or mention politically sensitive history, according to the archive, which makes its removal more revealing than an openly oppositional post would be.
The replacement notice cited violations of China's Cybersecurity Law, while the source says automated keyword scanning may also have contributed.
This is collateral censorship: content can disappear because it describes a sensitive place, even when its tone is neutral and its practical advice is simply to avoid the area in extreme heat.
The episode illustrates how security and censorship reinforce each other around symbolic sites, turning ordinary travel reporting into a potential compliance risk.
For platforms and users, the uncertainty is the point: broad rules and opaque enforcement encourage self-censorship beyond clearly prohibited political speech.

Open at China Digital Times Ask Claude ChatGPT

LandSpace makes China's first ground recovery of an orbital booster

# Solidot 5h

LandSpace's Zhuque-3 second flight delivered a satellite and landed its first stage in Gansu, giving China its first flight-proven orbital booster recovered on landing legs from land.

Also: SCMP

The 66-meter rocket used nine first-stage engines and landed six minutes after launch, after reaching orbit with its payload.
The milestone adds a land-recovery path to China's reusable-launch portfolio; the Long March 10B had previously demonstrated sea-based net recovery, according to the report.
A booster that survives an orbital mission is strategically valuable because reusability can reduce launch cost and increase cadence, but only if refurbishment is fast and reliable.
LandSpace's achievement puts its stainless-steel vehicle among a small group of flight-validated partially reusable orbital rockets.
The decisive follow-up is not the landing video but repeated launches, recovery of the same hardware and a cost structure competitive with expendable Chinese launchers.

Open at Solidot Ask Claude ChatGPT

China and the US still lack basic details for planned AI talks

# SCMP 3h

With Xi Jinping's planned US visit weeks away, Washington and Beijing have yet to settle the venue, participants or scope of a proposed AI dialogue, underscoring how difficult even narrow technical engagement has become.

Sources familiar with the talks say Washington has not decided which government body should lead, while the two sides have not agreed on whether technical experts or private-sector representatives will attend.
The uncertainty is politically important: AI talks are being treated simultaneously as strategic stability, technology policy and national-security business, with no obvious institutional owner.
A technical channel could cover frontier-model safety, compute controls, standards or crisis communication, but the scope remains unknown rather than merely undisclosed.
Beijing may prefer a state-to-state format that limits corporate exposure, while Washington must balance engagement with export-control and security concerns.
The timing around a presidential visit gives the talks symbolic value, but a meeting without working-level follow-through would produce little operational change.
The absence of basic arrangements this close to the visit suggests that political signaling is ahead of institutional preparation.

Open at SCMP Ask Claude ChatGPT

Chinese researchers entangle quantum memories across 420 kilometers

# Solidot 2h

A Chinese Academy of Sciences team reports entanglement between two cold-atom quantum memories across 420 kilometers of fiber, extending a key building block for metropolitan-scale quantum networks.

The team says it also exceeded the theoretical limit for direct, no-repeater entanglement distribution at distances above 230 kilometers.
Quantum-memory entanglement is relevant to repeater-based communications, distributed quantum computing and networked sensing because it links matter-based qubits rather than only transmitted photons.
The result builds on the group's reported progression from a 50-kilometer fiber link in 2020 to a three-node metropolitan network in Hefei in 2024.
The engineering challenge is not only distance but fidelity, synchronization and stable interfaces between memories, fiber and eventual repeater nodes.
China's repeated progress in this area shows a long-horizon state-supported research program rather than a single commercial product milestone.
A usable quantum network would require much higher operating reliability and repeaters; the report establishes a distance result, not a deployable internet.

Open at Solidot Ask Claude ChatGPT

China's space-computing boom hits a data-upload bottleneck

# GeekPark 9h

Chinese space-computing startups are pursuing orbital AI, but moving an 800-megabyte model to an in-orbit satellite can take nearly four weeks because low-Earth-orbit uplinks remain slow and intermittent.

The key distinction is between processing satellite-generated data in orbit and sending Earth-generated data upward for training or inference; the latter is the newer and harder business case.
The reported upload path operates around 1 Mbps and only during short ground-station windows, forcing a model into many fragments and making iteration painfully slow.
Downlink capacity has improved because satellites primarily send data to Earth, while high-speed uplink must overcome atmospheric distortion and was not designed for large files.
The bottleneck changes the commercial sequence: Chinese startups may first sell orbital storage or edge filtering while they develop a genuinely operable high-bandwidth uplink.
The story is a useful corrective to space-data-center hype because it identifies the missing system capability between an attractive architecture and a service customers can actually update.
If optical uplinks and launch cadence improve, orbital compute could become more than an edge-processing niche; until then, model refresh time is a fundamental constraint.

Open at GeekPark Ask Claude ChatGPT

China approves an environmental station near Scarborough Shoal

# SCMP 1h

Chinese authorities approved a marine observation station covering waters around disputed Scarborough Shoal, presenting a scientific-monitoring project that also strengthens Beijing's presence in a Philippine flashpoint.

The station was proposed by the Chinese Academy of Sciences' South China Sea Institute of Oceanology and approved on July 31, according to the institute.
Beijing's framing is environmental research and ecological observation, but the location gives Chinese institutions a continuing operational reason to gather data in contested waters.
Scientific infrastructure can serve a dual strategic role by improving maritime domain awareness, supporting legal claims and normalizing a state presence without being labeled a military installation.
For Manila, the concern is less the stated environmental mission than whether the station expands China's ability to monitor and operate around the shoal.
The project fits a broader Chinese pattern of using civilian agencies, research bodies and administrative infrastructure to reinforce claims in disputed maritime zones.
Its significance will depend on construction, access, sensor capabilities and whether other countries accept the station as a research facility rather than a sovereignty signal.

Open at SCMP Ask Claude ChatGPT

Geely founder Li Shufu gives way to veteran executive An Conghui

# Jiemian 8m

Geely Auto has replaced founder Li Shufu as chairman with longtime executive An Conghui, signaling a shift toward professional management as the company pivots fully to hybrids and overseas growth.

Li remains chairman of the larger Zhejiang Geely Holding Group and becomes Geely Auto's lifetime honorary chairman, separating group ownership from listed-company execution.
The reshuffle also moves Gui Shengyue to vice chairman and promotes Gan Jiayue to chief executive, suggesting a broader succession plan rather than a single title change.
Geely Auto reported first-half revenue of 173.6 billion yuan (roughly $24.5 billion), up 15%, while attributable profit fell 2% to 9.091 billion yuan (roughly $1.28 billion).
Management said the company will stop developing conventional gasoline vehicles and move its remaining combustion models toward intelligent hybrid technology.
The overseas sales target was raised to 920,000 vehicles, making execution outside China a central test for the new leadership.
The strategic question is whether a more professional structure can preserve Geely's multi-brand engineering discipline while reducing founder dependence.

Open at Jiemian Ask Claude ChatGPT

China's leadership signals span a senior funeral, Ecuador and AI policy

# Policy Sinocism 14h

A China policy digest highlights top leaders' public handling of Zhu Rongji's funeral, Xi's Ecuador meeting and domestic economic and technology policy, offering a compact view of Party-state priorities.

The funeral coverage shows the standard Party protocol for a former senior leader: current top officials attend, former leaders send wreaths or condolences and the event is tightly choreographed.
Xi's Ecuador meeting emphasized trade, energy, mining, infrastructure, finance, digital economy, green development, new energy and AI, showing how foreign economic policy is packaged as one development agenda.
The digest also tracks housing provident funds, consumption and lower-tier markets, indicating that domestic demand and household balance sheets remain central policy concerns.
Its value is synthesis: Chinese governance links elite protocol, overseas economic diplomacy and technology promotion within the same administrative narrative.
The source is an English-language China policy newsletter summarizing Chinese releases, so its interpretation should be separated from the official framing it quotes.

Open at Sinocism Ask Claude ChatGPT

Hong Kong's deadly heat is exposing gaps in climate policy

# Culture Initium 14h

Hong Kong endured record heat and unusually hot nights, while doctors and researchers warned that extreme temperature is becoming a larger public-health threat for outdoor workers and other vulnerable residents.

The city recorded 36.9 degrees Celsius on August 9 and a nighttime low of 30.2 degrees on August 12, according to the report.
The article describes a dense, humid city where heat stress compounds occupational risk, especially for construction and other outdoor workers who cannot simply retreat indoors.
Several suspected heat deaths during the heatwave make the issue more than a comfort or weather-alert story; it is a question of labor protection, housing and emergency care.
The coverage argues that Hong Kong's institutions have historically prioritized typhoons and rainstorms while heat lacks equally visible, enforceable adaptation measures.
As climate risk rises, cooling access, work-rest rules, neighborhood shelters and better mortality tracking will matter as much as temperature forecasts.
Hong Kong's experience is a useful case study for other Asian cities where density, humidity and an aging population amplify the cost of extreme heat.

Open at Initium Ask Claude ChatGPT

MiniMax loses a senior engineering leader during its agent push

# AI QbitAI 4h

MiniMax engineering chief Miao Yuhang has reportedly left as the company doubles down on coding, agents and long-context models, removing a visible bridge between model research and product engineering.

The source says his remit covered the M3.x model line, coding, agent infrastructure, audio and the Hailuo consumer application.
His role illustrates how Chinese model companies increasingly need leaders who can connect training systems, harnesses, tool use and end-user products rather than manage a single model team.
MiniMax has recently emphasized agent teams, coding and a million-token context, making the departure notable for organizational timing even though no reason beyond the report is established.
The next question is whether MiniMax can preserve continuity across model engineering and agent products without a similarly public technical coordinator.

Open at QbitAI Ask Claude ChatGPT

DeepSeek Harness becomes a community-built agent platform

# AI GeekPark 10h

DeepSeek Harness reportedly grew from a rough plugin-based release into a large community ecosystem within days, showing how open extension points can turn product incompleteness into a distribution advantage.

Also: V2EX

The source says the project reached about 149,000 GitHub stars, 15,000 forks, more than 5,100 plugins and over 3,500 authors five days after release; those figures are time-sensitive and source-reported.
Community developers built desktop wrappers, mobile interfaces, visual tools and skills that addressed complaints about installation, interface quality and missing multimodal features.
The architecture exposes adapters, tools, filesystem access, sandboxing, scheduling and even the agent loop as replaceable components rather than reserving a privileged core.
DeepSeek's internal engineering skills were also reportedly published, turning tacit code-review and cleanup practices into reusable, inspectable artifacts.
The tradeoff is governance: a highly modular agent can evolve quickly, but plugin permissions, supply-chain security, logging and rollback become part of the product's trust model.
The deeper signal is that a harness can be a platform even when the initial user experience is unfinished, provided the extension boundary is useful enough for others to repair it.

A community is turning DeepSeek Harness into a modular AI workbench V2EX

Open at GeekPark Ask Claude ChatGPT

Humanoid robot completes a full autonomous table-tennis match

# AI QbitAI 2h

At Beijing's World Robot Conference, SuperDynam KAI demonstrated what it calls the first complete autonomous humanoid table-tennis match, linking fast vision, trajectory prediction and whole-body control.

The SMASH 2.0 system must identify a fast, spinning ball, predict its trajectory and coordinate shoulder, elbow, wrist, torso and legs within millisecond-scale windows.
The company says the same control algorithm can run on different robot bodies, an important test of whether embodied policies are portable rather than tied to one machine.
KAI's world model was pretrained on millions of videos and is intended to generate and evaluate simulated interactions before deployment to hardware.
The demonstration is technically meaningful because table tennis exposes latency, contact dynamics and balance failures that slower pick-and-place demos can hide.

Open at QbitAI Ask Claude ChatGPT

A 5,100-yuan embodied-AI data stack targets the robot data bottleneck

# AI QbitAI 5h

51WORLD launched AperData at 5,100 yuan per set (roughly $720), combining first-person capture hardware with on-device quality checks and a platform for producing robot-training data.

The system shifts collection from expensive robot teleoperation to humans wearing cameras, inertial sensors and synchronized capture equipment, allowing more parallel data collection without consuming robot hours.
Its platform checks exposure, dropped frames, synchronization and task completeness near the source, filtering unusable segments before cloud upload.
The company claims physical trajectory consistency of 99% and more than tenfold data-production efficiency at equal cost; these are vendor claims, not independent benchmarks.
AperOne extends the stack into a five-stage application loop covering reconstruction, training, evaluation, deployment and operations in factories, mines and other sites.
The architectural bet is that embodied-AI economics depend as much on data yield and validation as on model quality, with edge-side quality control reducing storage, bandwidth and annotation waste.

Open at QbitAI Ask Claude ChatGPT

Tencent's recommendation contest tests a unified architecture for ads

# AI QbitAI 5h

A Tencent-backed KDD Cup challenge drew 13,913 participants from 52 countries to test whether sequence modeling and feature interactions can share one recommendation block without sacrificing millisecond latency.

The winning academic approach used an intent token as a persistent interest state, letting user profiles and behavior sequences update a shared representation through the network.
The industrial winner reportedly improved AUC by 0.0048 while reducing per-sample compute by 18%, a modest-looking gain that can matter at billion-scale advertising traffic.
The challenge exposes why recommendation systems cannot simply copy large-language-model scaling: ranking requests face far tighter latency and cost constraints than chat generation.
Moving from many specialized modules to a more uniform architecture could improve accelerator utilization and simplify maintenance, but the production tradeoff is still workload-specific.

Open at QbitAI Ask Claude ChatGPT

Chinese coast guard proposal imagines autonomous South China Sea outposts

# Policy SCMP 11h

A Chinese coast guard proposal envisions contested South China Sea islands defended by coordinated autonomous machines rather than soldiers, revealing how Beijing is thinking about drone-swarm asymmetry.

The proposal appears in a Chinese-language journal supervised by China State Shipbuilding Corporation, giving it institutional relevance even though it is not an announced deployment order.
The concept combines sensors, unmanned surface or aerial systems and autonomous coordination to make small, remote outposts harder to isolate or overwhelm.
Its strategic premise is that cheap drone swarms can threaten conventional garrisons, so a robotic defense layer could reduce personnel exposure while expanding surveillance and response capacity.
Autonomy would create serious command, identification and escalation problems in a crowded maritime environment where civilian and military activity overlap.
The proposal fits China's wider military-industrial focus on systems integration and mass-produced unmanned platforms rather than a single exquisite weapon.
The gap between a journal proposal and fielded capability is large; the important signal is the direction of planning, not evidence that robotic fortresses are imminent.

Open at SCMP Ask Claude ChatGPT

RSS research points to sample-efficient robot manipulation

# AI Leiphone 6h

Research presented at RSS 2026 shows Chinese and international teams using human videos, tactile alignment and synthetic trajectories to reduce the real-world data needed for dexterous robots.

A South China University of Technology team reportedly generated thousands of physically feasible bimanual training examples from one real demonstration by separating invariant coordination from object-specific adaptation.
Researchers from Shanghai AI Lab and Tsinghua reconstructed human hand-object interactions from monocular video and converted them into robot trajectories, targeting zero-shot deployment.
A Berkeley and Meta team aligned human and robot tactile observations in a shared latent space without paired datasets or manual labels, using under five minutes of human tactile data for a bulb-twisting transfer.
The common architectural move is to borrow structure from human demonstrations, geometry or touch instead of paying for millions of robot teleoperation hours.
These results point toward data engines that create physically constrained variations, but sim-to-real transfer, contact modeling and safety remain harder than the individual demonstrations suggest.

Open at Leiphone Ask Claude ChatGPT

A hit Iraq-war film reveals China's preferred view of intervention

# Culture SCMP 12h

A Chinese film about the Iraq war uses a Chinese chef's perspective to frame US intervention through ordinary civilian experience, extending Beijing's soft-power effort to shape how foreign audiences read its worldview.

The film avoids naming Iraq directly but depicts an invasion by American troops, allowing the political message to travel through allegory rather than an explicit state argument.
Centering a Chinese worker abroad gives the story an emotional bridge between China's own citizens and the consequences of great-power intervention.
The production is part of a broader cultural strategy in which historical and foreign conflicts are used to present China as an observer, victim or alternative source of order.
Its influence will depend on distribution and audience reception outside China; a state-aligned message does not automatically become effective soft power.
The film matters less as a literal account of the Iraq war than as evidence of how Chinese popular culture is being used to narrate American power and global instability.

Open at SCMP Ask Claude ChatGPT

A one-bit route to running giant models on small devices

# AI Leiphone 2h

At IJCAI 2026, researcher Haotong Qin presented post-training methods that compress large language models to one or two bits without full retraining, targeting the memory wall in edge AI.

The approach separates a small set of sensitive long-tail weights from the bulk of the model, allowing most weights to use one-bit representations while preserving key behavior.
The cited BiLLM method reportedly uses about half an hour of single-GPU processing rather than lengthy retraining; SqueezeLLM uses dynamic allocation in groups of 128 weights.
For a 7-billion-parameter model, moving from FP16 to four-bit representation can reduce roughly 14 gigabytes of weights to about one quarter of that size, cutting memory traffic as well as storage.
The remaining bottlenecks are activation and key-value-cache quantization, weak native hardware support below two bits and hidden losses in complex reasoning.
If these methods map cleanly onto mobile, automotive or robotic accelerators, they could shift more inference from centralized clouds to constrained devices.

Open at Leiphone Ask Claude ChatGPT

China's commercial space sector enters its scale-up phase

# Signal 36Kr 4h

A 36Kr research report says China's commercial space industry is moving from technical validation to scaled operations, driven by reusable rockets, satellite production and state-backed capital.

The report estimates the 2025 market at 2.83 trillion yuan (roughly $399 billion) and forecasts more than 3.5 trillion yuan (roughly $493 billion) in 2026, figures that should be treated as research-house estimates rather than official national accounts.
It cites 50 commercial launches in 2025, flexible satellite production lines and expanding low-Earth-orbit constellations as evidence of industrialization.
China is developing a mixed ecosystem in which state aerospace groups provide infrastructure and private companies pursue launch, satellite manufacturing and applications.
The report identifies bottlenecks in launch capacity, production efficiency, domestic components and orbital-spectrum access, all of which limit how quickly investment can become recurring service revenue.
It also points to in-orbit computing, servicing and satellite data applications as the next expansion areas, suggesting that the sector's value may shift from launch contracts to recurring infrastructure services.

Open at 36Kr Ask Claude ChatGPT

Taiwan's first domestic satellite launcher veers off course

# Policy SCMP 1h

Taiwan's first test of a locally developed satellite launcher failed after the vehicle veered off course, briefly triggering public fears that a missile had been fired.

The launch took place from a missile-testing base in southern Taiwan operated by the government-backed National Chung-Shan Institute of Science and Technology.
The incident highlights the dual-use nature of Taiwan's launch infrastructure, where civilian space ambitions and military testing facilities share a public and strategic context.
The account says nearby residents mistook the failed launch for an incoming missile, showing how limited public familiarity with the program can amplify security anxiety.
The test is an engineering setback, but it also gives Taiwan a visible reason to improve launch safety, public communication and separation between space and defense operations.

Open at SCMP Ask Claude ChatGPT

Baidu's AI revenue stays above half of the business

# Signal Leiphone 23h

Baidu reported 31.3 billion yuan (roughly $4.4 billion) in quarterly revenue, with AI accounting for half of general-business revenue for a second straight quarter as two US funds increased holdings.

Also: Jiemian

The company reported 25.2 billion yuan (roughly $3.5 billion) in general-business revenue, with AI contributing 50% according to the source.
The result gives investors a clearer measure of AI's weight inside Baidu, but it does not separate the economics of model services, cloud, advertising or autonomous driving.
The fund purchases are market sentiment, not proof that Baidu's search-to-AI transition is succeeding; they sit alongside the unresolved capital-allocation question raised in the separate analysis.

Baidu's AI business needs separation from search, argues Chinese analysis Jiemian

Open at Leiphone Ask Claude ChatGPT

Nokia plans to sharply shrink its China operations

# Culture Solidot 5h

Nokia reportedly plans to cut most China employees and close factories by year-end, retaining mainly after-sales service and unwinding a once broad local R&D and delivery footprint.

Nokia had about 7,200 employees in mainland China and Hong Kong at the end of 2025, according to its annual report.
The company's China presence historically included research, marketing, services, global delivery and supply operations, so the move suggests a retreat from local industrial depth rather than a simple sales-office cut.
The reported plan reflects the pressure foreign network-equipment companies face as Chinese operators favor domestic suppliers and global firms consolidate supply chains.
For Chinese engineers and suppliers, the important signal is the loss of an international R&D and manufacturing platform that had operated in Beijing since 1995.

Open at Solidot Ask Claude ChatGPT

Digital-signal-processing pioneer B. T. Liu dies at 91

# Culture Solidot 4h

Engineer and educator B. T. Liu, whose work helped move audio, image and video processing from analog to digital systems, has died at 91.

Liu was born in Shanghai, later studied and worked in Taiwan and the United States, and joined Princeton University in 1962.
His 1976 textbook on digital signal processing became a standard engineering reference, helping define a field now embedded in communications, streaming and low-power devices.
The story is a reminder that much of today's infrastructure rests on mathematical and systems work that predates the current AI cycle by decades.

Open at Solidot Ask Claude ChatGPT

Hong Kong exchange posts record first-half profit as China trading surges

# Signal Jiemian 14m

Hong Kong Exchanges and Clearing reported record first-half profit of HK$10.568 billion (roughly $1.49 billion), as cash, derivatives and mainland link volumes all reached new highs.

Revenue and other income rose 19% to HK$16.702 billion (roughly $2.35 billion), while attributable profit rose 24%.
Average daily cash-market turnover reached HK$283 billion (roughly $39.9 billion), up 18%, and exchange-traded product turnover rose 28% to HK$48.4 billion (roughly $6.8 billion).
The exchange attributed the rebound to renewed global interest in Chinese technology and artificial-intelligence stocks, active new listings and participation from both mainland and international investors.
The performance makes Hong Kong a useful barometer for whether enthusiasm for Chinese assets is broadening beyond a small group of technology names.
The signal is cyclical as well as structural: record volumes improve exchange economics, but sustained listings and cross-border participation are needed to prove a durable market revival.

Open at Jiemian Ask Claude ChatGPT

Mongolia apologizes after dispute with Chinese-backed oil explorer

# Policy SCMP 6h

A Mongolian official apologized after China's embassy objected to protesters confronting a Chinese-invested oil exploration operation, exposing the political risk around Chinese projects abroad.

The protest occurred at the Kherlen Tohoi exploration area, where Mongolia Foison Energy was working, according to Mongolian media cited by SCMP.
Chinese diplomatic pressure turned a local environmental or land-use dispute into a bilateral issue, with the embassy describing the protesters as harassing the camp.
The incident shows that Chinese investment does not automatically translate into local legitimacy; project security and community relations can become matters of state diplomacy.
For Chinese firms operating overseas, the operational lesson is that labor, land and environmental grievances can rapidly acquire a China-versus-host-country frame.

Open at SCMP Ask Claude ChatGPT

A Chinese physics program explores low-power materials with memory

# Culture QbitAI 6h

A Sohu physics program brought a Chinese science communicator together with University of Pennsylvania physicist Andrea Liu to discuss jamming, soft matter and materials that could compute with very low power.

The discussion uses foams, sand and disordered particles to explain jamming, where a system becomes rigid after particles cross a critical density.
The scientific idea is relevant to unconventional computing because collective material states could encode memory or learning without the energy costs of repeatedly moving digital bits.
This is a public-science conversation, not a reported device breakthrough; the potential applications remain exploratory.

Open at QbitAI Ask Claude ChatGPT

Power failure on US destroyer raises questions about reserve capacity

# Policy SCMP 5h

An analyst told SCMP that a US destroyer's loss of power in the South China Sea may reflect strained reserve capacity as Washington's Iran operation continues, though the explanation is unproven.

The incident matters to China because the South China Sea is a high-tempo operating environment where maintenance failures can affect deterrence and signaling.
The source quotes a former US Navy captain calling the event highly unusual and suggesting multiple equipment or process failures would have been required.
This is analysis rather than an official investigation, so it does not establish a systemic readiness problem or attribute a cause.
If repeated failures appear across deployments, the issue could affect how Beijing assesses US operational endurance; a single breakdown is a much weaker signal.

Open at SCMP Ask Claude ChatGPT

EVE Energy doubles first-half profit on battery shipments

# Signal IT Home 1h

EVE Energy reported first-half revenue of 45.691 billion yuan (roughly $6.4 billion) and profit of 3.301 billion yuan (roughly $465 million), driven by strong power and storage-battery volumes.

Power-battery shipments rose 66.47% to 35.76 gigawatt-hours, while energy-storage shipments increased 54.88% to 44.46 gigawatt-hours.
The company says its 46-series cylindrical cell has entered volume production and is the launch battery supplier for BMW's next-generation vehicles.
Malaysia has begun producing consumer batteries, giving EVE a more global manufacturing footprint as Chinese battery makers diversify production.
Operating cash flow was negative 388 million yuan (roughly $55 million), so the profit surge still needs to be weighed against working-capital and expansion demands.

Open at IT Home Ask Claude ChatGPT

Robam is building an AI operating system for the kitchen

# Culture 36Kr 8h

Chinese appliance maker Robam is positioning an AI cooking agent as a coordinator for ovens, burners and other devices, using decades of cooking data to move beyond connected-appliance features.

The proposed architecture pairs a central cooking agent with device-level agents that can respond locally while following a shared plan and safety state.
Robam says it has 50 million household users, a database of more than 5,000 cooking curves and 827 AI-enabled product models; these are company-reported figures.
The kitchen is a demanding embodied-AI environment because heat, smoke, liquid, ingredients and timing change continuously, making sensing and closed-loop control more important than a voice interface.
The business shift is from selling hardware once to combining devices, software, recipes, usage data and ongoing service, a model Chinese appliance companies are increasingly pursuing.
The claims remain largely promotional, and the real test is whether the system can improve cooking outcomes safely across diverse homes rather than only coordinate appliances in demonstrations.

Open at 36Kr Ask Claude ChatGPT

China's robot IPO and industrial training bets dominate a tech roundup

# Leiphone 12h

A Chinese technology roundup links Unitree's IPO, Cambricon employee equity, Gree's company-run vocational school and new robot supply-chain bets, showing how robotics is spreading through finance, labor and education.

Also: 36Kr

The roundup says 124 Cambricon employees will receive shares worth about 690 million yuan (roughly $97 million) at the cited market price, illustrating how China's AI-chip boom is affecting technical labor incentives.
Gree's new school will enroll 300 teenagers and connect training directly to the company and its suppliers, reviving a more explicit school-to-factory employment model.
Its Unitree coverage emphasizes that internet companies, state-backed funds and industrial investors have accumulated stakes around the robot maker's listing.
As a roundup, the item is evidence of Chinese media attention and several related signals, not a single independently developed story.

China's morning tech roundup mixes robot finance with AI workplace tools 36Kr

Open at Leiphone Ask Claude ChatGPT

Kuaishou expects AI-generated comic series to reach 500,000 in 2026

# Culture Jiemian 25m

Kuaishou says AI is lowering content-production costs and expects its animated short-series output to rise from 60,000 titles in 2025 to 500,000 this year, reshaping China's mobile entertainment supply.

The forecast comes from Kuaishou chief executive Cheng Yixiao during the company's second-quarter earnings call, so it is a management projection rather than audited output.
The implied expansion is more than an eightfold increase in series count, showing how quickly inexpensive AI production can flood a short-video platform with serialized content.
China's short-video companies are not treating generative AI only as a creator tool; they are also using it to expand the catalog that keeps users inside their recommendation systems.

Open at Jiemian Ask Claude ChatGPT

Ping An Good Doctor grows enterprise health management revenue 65%

# Signal 36Kr 1d

Ping An Good Doctor's first-half revenue was nearly flat, but enterprise health-management revenue rose 65% to 28.7% of sales as AI helped expand its medical service network.

The company reported revenue of 2.484 billion yuan (roughly $350 million) and attributable profit of 219 million yuan (roughly $31 million), up 63.5%.
It says more than 7,700 paying enterprise customers used its services in the past year, while the business generated 1.79 billion yuan (roughly $252 million) in gross merchandise value.
Ping An says its medical models cover more than 11,300 diseases with diagnostic accuracy near 96%, but those are company-reported metrics and not a substitute for clinical validation.
The business model is notable because AI is being used to improve a hybrid online, hospital, home-care and employer network rather than sold as a standalone chatbot.

Open at 36Kr 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	1/20
QbitAI	200	10/10
36Kr	200	7/30
Solidot	200	12/20
Leiphone	200	16/20
GeekPark	200	4/30
IT Home	200	12/60	+47 over cap
SCMP	200	10/50	+9 over cap
Initium	200	3/15
Jiemian	200	8/30	+2 over cap
China Digital Times	200	1/15
The Wire China	200	0/50
Techmeme	200	0/15
Hacker News	200	1/20
V2EX	200	8/50	+29 over cap

Unitree took $48 billion out of a falling market, and its earliest backer says most of this sector dies

Unitree up 460%, Baidu down 12%

LandSpace landed a steel booster on legs

Nobody knows who runs the AI talks

DeepSeek's harness now ships whole agents

A Qianmen travelogue got deleted anyway

Threads we are pulling