China Desk — 18 Aug 2026, 01:45 UTC

Two Chinese humanoids complete an autonomous 11-point table-tennis game

# QbitAI 16h

A Hong Kong University team says two humanoid robots played a full 11-point game without remote control or human feed, using its SMASH 2.0 system.

SMASH links vision, trajectory prediction, action planning and whole-body control rather than asking one large model to handle the entire loop.
The 2.0 release expands the reachable ball region to short and long shots and adds autonomous serving, both necessary for a real game rather than a sequence of assisted returns.
The team trained with human motion-capture data collected for four to eight hours a day over one to two months, while preserving successful and failed robot rallies for later reinforcement learning.
The system still uses preset strategies and does not learn an opponent's style online; rotation handling is identified as a later challenge.
A complete game is a stronger autonomy test than keeping a rally alive because each robot must pursue winning behavior, not merely accept a human-fed ball.
The demonstration is a valuable control benchmark, but it does not yet show competitive play against skilled human athletes or broad transfer beyond table tennis.

Chinese humanoids play a full autonomous table-tennis match QbitAI

Open at QbitAI Ask Claude ChatGPT

China's EV boom grew through local alliances, not just central planning

# ChinaTalk 15h

A ChinaTalk discussion argues that entrepreneurs, local officials, foreign capital and informal networks built China's EV industry, while leaving the country with serious overcapacity.

The account traces the industry from roughly 130 small automakers in the 1990s to globally competitive firms such as Chery and Geely through local-government alliances and private capital.
Its central correction to simple industrial-policy narratives is that local officials and entrepreneurs often worked around state-owned incumbents and central regulation rather than following a single national blueprint.
The same decentralized growth model helped create excessive capacity, making exports and overseas factories an outlet for domestic competition.
The episode treats personal networks and elite politics as mechanisms for deciding which firms receive access to land, finance, technology and market opportunities.
The proposed lesson for robotics and AI is that local experimentation can build champions quickly, but coordination failures and overinvestment may travel with the model.
This is analysis and a podcast interview, not a new policy announcement, but it is a useful map of how Chinese industrial development actually gets assembled.

Open at ChinaTalk Ask Claude ChatGPT

Weak July data, a State Council plenary and Xi's Washington visit

# Sinocism 3h

A China-focused newsletter flags broadening domestic-demand weakness alongside a State Council plenary, a Jiang Zemin commemoration and a possible Xi visit to Washington.

The roundup says service retail sales rose 5.0% in January through July, while goods retail grew 3.9 percentage points more slowly and online retail rose 4.8%.
It highlights faster growth in electronic-circuit manufacturing investment, reported at 57.7%, and integrated-circuit investment at 11.5%, as high-tech and digital manufacturing supplied about half of industrial value-added growth.
The juxtaposition captures Beijing's current growth model: weak household demand is being offset politically and financially by investment in strategic manufacturing.
The newsletter also places a State Council plenary and a possible Xi visit to Washington in the same information frame, linking domestic economic management with external bargaining.
The supplied excerpt is a roundup lead, not a full account of the policy decisions or the diplomatic agenda, so the institutional implications remain unclear.
For executives, the signal is a policy mix that continues to privilege industrial upgrading even as consumption momentum softens.

Open at Sinocism Ask Claude ChatGPT

DeepSeek makes the entire coding-agent harness modular

# Leiphone 16h

DeepSeek's developer preview treats the model adapter, tool registry, session log and agent loop as replaceable plugins, enabling a self-evolving runtime with rollback mechanisms.

The Cordis framework records reversible side effects so a removed component can undo listeners, timers and resource changes instead of leaving the runtime contaminated.
Its reactive dependency model lets components activate, deactivate or remain unchanged as providers appear, disappear or are replaced.
This is a deeper architectural move than adding tools around a fixed agent: the execution loop itself becomes mutable, allowing a harness to inspect and change its own runtime.
The design addresses the obvious failure mode of self-modification by making changes reversible and dependencies explicit, though the preview does not prove safe autonomous evolution.
The project reportedly passed 45,000 GitHub stars shortly after release, a strong developer-interest signal but not evidence of production reliability.
For enterprise architects, the key question is whether plugin-level mutability can coexist with deterministic policy enforcement, audit trails and reproducible execution.
If it works, DeepSeek is proposing a runtime substrate for agents rather than another coding assistant, with implications for model portability and tool governance.

Open at Leiphone Ask Claude ChatGPT

China's embodied-AI industry shifts from doing tasks to closing workflows

# Leiphone 22h

A Chinese robotics company is testing a layered Agentic OS across robots, an AGV and a robot dog in an automotive logistics line, measuring whether the full task completes rather than whether one action works.

The company reports that domestic embodied-AI financing reached 93.5 billion yuan, roughly $13.2 billion, in the first half, five times the year-earlier figure, while actual robot shipments remain far smaller than the funding frenzy.
Its architecture separates high-level planning from an on-site runtime for state monitoring and recovery, leaving real-time control and safety to device-specific models.
That separation addresses two hard industrial constraints: cloud-model latency and the cost of hallucinated actions around people and machinery.
The article argues that factories must also redesign processes around what robots can execute reliably, rather than forcing robots into workflows optimized for human bodies.
A task log that records failure, recovery and continuation is treated as more valuable than another training dataset because deployment is effectively a test set for the system.
The strategic shift is from equipment-centric automation to software-defined task orchestration, where completion and verification become the unit of value.
The approach remains an early industrial trial, but it identifies the systems-engineering bottleneck that model demos routinely obscure.

Open at Leiphone Ask Claude ChatGPT

Alibaba's Qwen passes 3 billion downloads in six months

# Solidot 10h

A Hugging Face report says Alibaba's open-weight Qwen models drew more than 3 billion global downloads in six months, far ahead of Google's 418 million and Meta's 227 million.

The reported community footprint is substantial: 151,448 derivative models are based on Qwen, compared with 82,506 for Google, according to the supplied figures.
The report says Chinese labs released high-performing open-weight models almost monthly in 2026, with some parameter counts reaching the hundreds of billions or trillions.
Open weights are not the same as unrestricted commercial use: the article says many Chinese models carry noncommercial limits, and some newer Qwen releases require revenue sharing.
Qwen's lead therefore comes from distribution and ecosystem reuse as much as benchmark performance, creating a developer network that can compound even when licenses remain commercially constrained.
The result suggests Chinese labs are using open distribution as an international adoption strategy despite restricted access to leading chips.
The figures come from a cited Hugging Face report and should be checked against its methodology, especially the meaning of a download and the treatment of derivative models.

Open at Solidot Ask Claude ChatGPT

Pinduoduo's rural delivery network turns village logistics into a farm input

# Leiphone 22h

A Pinduoduo pilot in Henan lets delivery trucks carry parcels into villages and pick up garlic on the return trip, cutting a seller's shipping cost from about 7 yuan to 3 yuan per five-jin package.

The model removes the empty-return problem that previously made village collection uneconomic, allowing local sellers to ship produce without first hauling it to a county hub.
The reported pilot handles 10,000 to 12,000 incoming parcels a day and about 2,000 outgoing parcels, most of them local garlic.
The infrastructure changes who can participate in e-commerce: small farmers and village sellers gain a route to national buyers without building their own transport network.
The village pickup point also creates foot traffic for local shops, showing that logistics infrastructure can reshape rural retail and employment as well as shipping costs.
The article is platform-sponsored and relies heavily on participant testimony, so the durability of the savings and the program's subsidy economics remain unclear.
Strategically, this is a Chinese platform using logistics density to make agricultural supply more legible and liquid, rather than merely offering another online storefront.

Open at Leiphone Ask Claude ChatGPT

Waymo imported 3,200 Chinese Zeekr cars despite US tariffs

# Solidot 17h

Waymo has imported 3,200 Zeekr vehicles since 2024 because, even after a reported 127.5% tariff, they remain cheaper to convert than its Jaguar-based alternative.

The reported Zeekr vehicle cost rises from about $38,000 to $86,500 after tariffs, and adds roughly $25,000 for autonomous hardware and software, still keeping the total near $110,000.
Waymo's Jaguar I-Pace platform reportedly exceeds $200,000 after modification, showing that Chinese vehicle economics can survive severe trade barriers when the alternative is a high-cost domestic platform.
The episode exposes a gap between industrial policy and deployment economics: tariffs can restrict consumer imports without eliminating a Chinese vehicle's value in a specialized fleet.
It also shows why autonomous fleets care about body-in-white cost, packaging and retrofitability as much as battery range or consumer brand.
The figures come from the supplied report and do not establish whether future US rules will permit the same sourcing pattern.

Open at Solidot Ask Claude ChatGPT

Honor bet a $1,410 gimbal phone on a new AI-device category

# 36Kr 1h

Honor says its Robot Phone combines a five-year-rated motorized gimbal, ARRI imaging workflows and a system-level agent in a bid to create a second growth curve during a handset downturn.

Also: IT Home

The mechanical package contains more than 100 precision parts, 68 board-to-gimbal wires and 19 coaxial lines inside a 9.59mm handset weighing 248 grams.
Honor says an early hinge design lasted about one year; after testing 18 material options, it raised the target design life to five years, highlighting the reliability gap between a camera accessory and a daily phone.
The camera stack supports ARRI LogC3, LUT workflows, focus and exposure locks, and subject tracking, positioning the device between a smartphone and a stabilized camera.
The Yoyo agent is intended to combine image, sound, gestures and long-running commands, expressing Honor's view that the phone should become a carrier for agents rather than an app container.
The article reports more than 400,000 reservations and a rapid sellout, but those are company-reported launch signals rather than sustained demand.
Magic9 is expected to inherit the phone's AI and imaging technology, making this a platform experiment rather than a one-off luxury handset.

Open at 36Kr Ask Claude ChatGPT

Indonesia's top Chinese 3C seller faces a currency and margin problem

# 36Kr 22h

Chinese accessories company Vivan says Indonesian revenue reached 1.217 billion yuan, roughly $171 million, in 2025, but exchange losses and platform fees are eroding profits before its Hong Kong listing.

The company gets 94% of revenue from Indonesia and relies on more than 44,000 distributors, 42 local offices and nine warehouses, making its physical network more valuable than any single online channel.
Revenue rose 22% in the first half of 2026 while operating profit was nearly flat; a reported 179 million yuan, roughly $25 million, currency loss consumed about 72% of first-half net profit.
The business buys heavily in China but prices in Indonesian rupiah, so currency depreciation hits both reported revenue and the cost base; the company says it moved to yuan settlement and uses forward contracts.
Marketplace commissions also rose sharply, showing that a company can be first in a fragmented national category while still losing economics to platform distribution.
The company's reported 2.2% share of Indonesian 3C accessories indicates that first place does not mean market power; the local distribution footprint is the harder asset to reproduce.
The IPO case is a useful warning for Chinese companies going abroad: localization creates reach, but it also imports local-currency, channel and governance risk.

Open at 36Kr Ask Claude ChatGPT

HelixWorld adds spatial audio to real-time interactive world models

# AI QbitAI 18h

A Hong Kong-linked research team released HelixWorld, which jointly generates interactive video at 24 frames per second and 48kHz stereo audio in response to user movement.

The system uses a shared Transformer to generate video and audio together instead of adding a soundtrack after rendering, so camera motion can change both visual perspective and sound fields.
Its data pipeline combines first-person real-world video with game-engine data, where geometry, material, source position and listener orientation are known and can supply aligned spatial labels.
The team reports million-scale training clips filtered for frame-level audio-visual alignment, source direction and temporal correspondence between events and sound energy.
The model keeps separate audio and video latent representations but lets them exchange information inside one generation process, a useful compromise between modality-specific encoding and joint dynamics.
The planned open release could make interactive audio-world modeling easier to reproduce, although the supplied text does not provide comparative latency, quality or hardware requirements.
The technical challenge is not adding sound; it is preserving causal, spatially correct sound as the user changes viewpoint and action in a continuously generated world.

Open at QbitAI Ask Claude ChatGPT

How DJI rebuilt the technical stack for 8K 360-degree cameras

# Analysis GeekPark 30m

DJI's Osmo 360 line uses a purpose-built square sensor and lessons from drones and gimbals to push native 8K capture beyond the usual phone-derived architecture.

DJI says conventional 360-degree cameras waste the sides of rectangular phone sensors; its 1/1.1-inch square CMOS uses the circular image field more efficiently and produces native 8K from two 4K views.
The company reportedly spent two years co-developing the sensor and then pushed the image pipeline to 8K at 50 frames per second through overclocking and stress testing.
The article's deeper point is architectural: DJI reused drone stitching, spatial reconstruction and stabilization expertise instead of assembling commodity camera components.
Its later 360-degree drone work exposed harder cases such as parallax, seam visibility and high-speed motion, turning an unusual camera into a platform for algorithmic validation.
This is a Chinese hardware company moving down the stack from product integration toward custom sensors and imaging compute, a pattern worth watching beyond cameras.

Open at GeekPark Ask Claude ChatGPT

A Chinese startup tests full-body intelligence with a robot racing a kart

# AI Leiphone 16h

Symbiosis Robotics showed a bipedal humanoid driving a go-kart, using one continuous task to test perception, balance, hand-foot coordination and force control.

The company describes the kart not as a product but as a stress test for an end-to-end vision-to-action model that must coordinate steering, pedals and balance in a confined cockpit.
That is a harder benchmark than walking or fixed-position grasping because perception, motion control and hardware constraints must remain coupled over a continuous sequence.
The architecture is intended to reduce information loss between modular systems, but the source explicitly says end-to-end design does not prove general-purpose capability.
The team says its members have backgrounds in Chinese research institutes, universities and major technology companies, with prior work in VLA models, whole-body control and cross-platform learning.
The video is an early validation artifact: model details, test conditions and reproducible success rates have not yet been disclosed.
Its strategic value is the direction of travel, from demonstrating motion primitives to measuring whether an embodied model can complete a physically entangled task.

Open at Leiphone Ask Claude ChatGPT

A Peking University professor bets on latent world-action models

# Analysis Leiphone 22h

BeingBeyond says it has collected more than 500,000 hours of first-person human video and is building an embodied model that predicts actions and future states in latent space rather than generating pixels.

The company's latent world-action route is reported to cost about 1% as much to train as video-generation approaches and to support real-time control, at the cost of being much harder to show in a demo.
Its Being-H0.8 model adds touch to vision, action and future-state prediction in one latent space, aiming to represent how contact changes the world.
Founder Lu Zongqing argues that embodied AI lacks a settled equivalent of the Transformer and next-token prediction, so scaling data before the control paradigm converges may be premature.
The 500,000-hour first-person dataset is a significant bet on human video as the most scalable source of embodied experience, but video alone does not supply all the action, force and embodiment labels a robot needs.
The interview also highlights a funding problem: a three-to-five-year technical horizon is mismatched with investors seeking short-cycle valuation events.
For executives, the useful distinction is between a world model that renders plausible pixels and one that predicts state transitions useful for control; the field has not yet established which route will win.

Open at Leiphone Ask Claude ChatGPT

Pentagon orders 30 universities to audit research ties with China

# Policy SCMP 10h

The Pentagon ordered 30 US universities to review foreign research and financial ties, including Chinese partners, or risk losing future federal funding.

The directive covers collaborations with foreign entities of concern and organizations linked to former Confucius Institutes, according to the supplied report.
This turns research-security screening into a funding-eligibility condition, giving universities a strong incentive to widen audits beyond classified work.
The China impact is practical: ordinary academic partnerships may face more legal review, slower contracting and higher compliance costs even when no sensitive technology is involved.
The supplied text does not name the universities or define the final threshold for a prohibited relationship.
For technical leaders, the policy points toward a more fragmented research ecosystem in which joint labs, visiting scholars and shared datasets require institutional provenance checks.

Open at SCMP Ask Claude ChatGPT

China's robotics expo is likely to reveal engineering, not magic

# Analysis GeekPark 22h

A GeekPark preview argues that the World Robot Conference should be judged by autonomy, data tooling, force sensing, non-humanoid forms and continuous operation rather than stage demos.

The article notes Unitree's planned listing at an implied value of about 61 billion yuan, roughly $8.6 billion, against reported 2025 humanoid shipments of more than 5,500 units.
That valuation-versus-volume gap captures the industry's current condition: capital expectations are converging faster than data, hardware cost and reliable deployment.
The five questions are practical benchmarks: whether non-humanoids can be more useful, whether long-horizon tasks work, whether data collection is becoming a hardware business, whether key components are improving and when economics can approach EV scale.
The preview says the sector has not converged on a body, dataset or algorithm, but is progressing through incremental improvements in motors, fingers, tactile sensing and uptime.
For technical leaders, the useful advice is to inspect two-hour continuous operation and recovery behavior rather than be impressed by a single successful motion.

Open at GeekPark Ask Claude ChatGPT

A study finds Chinese Wikipedia deletion decisions highly concentrated

# Culture Solidot 12h

A study of 47,030 Chinese Wikipedia deletion cases from 2020 to 2024 found that a small group of administrators made most binding decisions, with one handling 89.9% in 2023.

The study reports that six administrators made 85% to 100% of decisions, while one administrator handled 6,312 cases in 2023.
Although simplified-Chinese users vastly outnumber traditional-Chinese users, the study says traditional-Chinese editors are 1.2 to 1.8 times more numerous, exposing a striking participation imbalance.
Deletion debates most often cited insufficient notability, and more than a quarter of deletions reportedly lacked a clear reason.
The concentration creates a governance problem familiar to online communities: nominally open deliberation can still produce highly centralized practical authority.
The result is a useful counterpoint to assumptions that platform governance is determined by the largest language population; participation quality and administrator availability matter more than raw readership.

Open at Solidot Ask Claude ChatGPT

Beijing is building an independent theory of international relations

# Policy Pekingnology 13h

A Chinese academic essay describes the Party-backed push for an autonomous knowledge system and argues that international-relations theory should not be treated as synonymous with Western theory.

The initiative traces back to Xi Jinping's calls for Chinese-characteristics philosophy and social sciences and is being organized through Party journals and universities.
The essay surveys relational theory, moral realism, world-order theory and symbiosis theory before calling for more systematic use of traditional thought and China's diplomatic experience.
This is not simply an academic branding exercise: it is part of a state effort to align social-science disciplines with national political experience and strategic autonomy.
For foreign companies and researchers, the trend suggests that Chinese policy concepts may increasingly be justified through domestically developed theoretical vocabularies rather than imported Western frameworks.

Open at Pekingnology Ask Claude ChatGPT

China's rocket failures expose a shortage of room for failure

# Analysis SCMP 4h

An SCMP analysis argues that rapid unofficial explanations after a Long March 7A failure reveal how difficult it is for China's space sector to discuss failure openly.

CASC officially described the August 10 launch only as a flight anomaly, while trade reports quickly pointed toward an engine; the cause remained unconfirmed in the supplied text.
The speed of consensus is the story: an ecosystem that rewards successful launches and controlled messaging may leave little room for engineers to publish contradictory hypotheses.
That matters technically because reusable or rapidly iterated launch systems need failure data to move through design-test-redesign loops rather than treating each anomaly as reputational damage.
China's state-linked space structure can mobilize capital and manufacturing quickly, but the same hierarchy may constrain independent debugging and public accountability.
The article does not establish the failure cause or prove that disclosure rules caused it; it raises a governance question about how learning is organized.

Open at SCMP Ask Claude ChatGPT

A Zhejiang University study says restraint improves VLM fine-tuning

# AI Leiphone 21h

Researchers propose an adaptive asymmetric adapter that often leaves the image encoder mostly untouched to preserve out-of-distribution generalization in CLIP-like models.

Experiments across 11 vision datasets reportedly found that text-branch tuning usually helps more than image-branch tuning, while aggressive image tuning can damage OOD performance.
The proposed adapter uses one shared down-projection and multiple expert up-projections, with a router that chooses how much visual adaptation each input needs.
A confidence-based gate can effectively return uncertain inputs toward the pretrained visual backbone instead of forcing every sample through a specialized adapter.
This is a useful engineering correction to the assumption that more image-side tuning always improves a vision-language model; in production, test-set gains can hide a brittle visual representation.
The supplied article is a technical explainer of a paper, so the generalization claims still need review of the paper's splits, baselines and compute budget.

Open at Leiphone Ask Claude ChatGPT

The real AI bottleneck may be verification, not model size

# Analysis SCMP 13h

A China-focused analysis argues that excitement around Kimi K3 and Qwen3.8 ran ahead of auditable benchmarks, exposing a verification gap in the global AI race.

The article says markets reacted to the Chinese model releases before full benchmark tables, model cards and licenses were publicly auditable, contributing to a roughly $3 trillion fall in chip-stock value.
That framing shifts attention from who has the largest model to whether outside developers can reproduce claims, inspect licenses and evaluate deployment costs.
For Chinese labs, rapid releases and open-weight distribution can generate global attention quickly, but missing documentation weakens trust with enterprise buyers and researchers.
The issue is not uniquely Chinese: the broader industry incentive is to announce capability first and disclose evaluation details later.
The practical implication is that model cards, reproducible tests and clear commercial terms are becoming infrastructure for adoption, not optional communications material.

Open at SCMP Ask Claude ChatGPT

Ecuador's president courts Beijing for shrimp and mining investment

# Policy SCMP 4h

Ecuadorian President Daniel Noboa is visiting China to seek the reopening of shrimp processors and new energy and mining investment while preserving ties with Washington.

Beijing says Noboa will meet Xi Jinping, Premier Li Qiang and the country's top legislator during an eight-day state visit.
The immediate economic agenda combines market access for 14 suspended shrimp processors with capital for energy and mining, showing how China uses trade remedies and investment as linked diplomatic tools.
Ecuador's attempt to avoid choosing between Beijing and Washington reflects the bargaining space available to middle powers amid US-China competition.
The supplied extract does not say whether the shrimp suspensions will be lifted or what investment agreements will be signed.
For Chinese companies, the visit may create openings in resource and infrastructure projects, but political balancing in Quito will remain a constraint.

Open at SCMP Ask Claude ChatGPT

HiDream releases an interactive world model built for stable physics

# AI Leiphone 16h

Chinese startup HiDream says its O1 World model can generate and edit explorable scenes while maintaining spatial and physical consistency across camera movement.

The system combines a native multimodal architecture with persistent memory and test-time training to reduce scene drift during long interactions.
The company says it can generate worlds from text or images, support first- and third-person navigation and edit actions such as grasping, jumping and rainfall in real time.
On the WBench navigation split, it reportedly scored 80.9 overall, 73.3 on physics and 88.0 on consistency across 289 cases and 1,058 interaction rounds.
Those figures come from a company announcement about a third-party benchmark, and the supplied text does not provide the comparison set or reproducible evaluation code.
The technical target is important for robotics and simulation: stable geometry, object permanence and causal responses are prerequisites for a world model that can support planning rather than just render attractive frames.

Open at Leiphone Ask Claude ChatGPT

3D Gaussian splatting moves from visual demos toward robot infrastructure

# AI Leiphone 21h

A CVPR 2026 review highlights feed-forward reconstruction, adaptive sparsity, collision modeling and noise-resistant mapping as 3D Gaussian splatting's path into embodied AI.

EcoSplat reportedly lets a model choose a top-K set of Gaussian primitives for a device's compute budget, preserving 24.72dB PSNR with only about 5% of its original points on RealEstate10K.
SparseSplat allocates fewer primitives to low-information surfaces and more to detailed regions, reaching comparable quality with 150,000 rather than 6.88 million Gaussians on a cited dataset.
At extreme compression, the review reports rendering rates of 600 to more than 1,100 frames per second, pointing toward mobile and robot-side deployment rather than offline visualization.
The broader shift is from photorealistic reconstruction as an end product to a controllable scene representation that can support simulation, collision reasoning and policy training.
Several highlighted results involve Chinese universities and companies, but the review is a survey of conference work rather than a single Chinese commercial milestone.

Open at Leiphone Ask Claude ChatGPT

The Arctic route becomes a strategic option for China and Russia

# Policy SCMP 10h

Chinese shipping through Russia's Northern Sea Route is gaining importance as Red Sea disruption makes the shorter Arctic path more commercially relevant.

The route spans about 5,600 kilometers along Russia's Arctic coast and can shorten Europe-Asia voyages, but navigation remains dependent on ice conditions and seasonal access.
The commercial case is being strengthened by conflict-driven risk around the Suez route, not only by a desire to explore a new lane.
China's participation gives Russia a customer and partner for Arctic infrastructure while giving Chinese shippers another option around maritime chokepoints.
The environmental tradeoff is material: more traffic can accelerate local ice disruption, while warming already makes Arctic ice thinner and less predictable.
The article frames the route as strategically important, but the supplied text does not establish that it is yet a dependable substitute for Suez.

Open at SCMP Ask Claude ChatGPT

Taiwan and a US startup will develop autonomous underwater drones

# Policy SCMP 12h

Taiwan's government-funded weapons institute signed an agreement with US startup Vatn Systems to develop autonomous underwater vehicles amid pressure from Beijing.

The deal expands Taiwan's asymmetric-defense emphasis toward unmanned underwater systems, where relatively inexpensive platforms can complicate maritime planning.
Because the work runs through Taiwan's top weapons developer, it is more than a private-sector trial and could feed into state procurement or indigenous defense programs.
The China dimension is direct: the project is explicitly framed as a response to military pressure from Beijing.
The supplied extract does not specify vehicle range, autonomy architecture, production numbers or delivery dates, so the operational value remains unclear.
For US-China technology watchers, the agreement adds another channel through which American dual-use robotics expertise can enter Taiwan's defense ecosystem.

Open at SCMP Ask Claude ChatGPT

Fake mobile agents leave Xi'an residents with surprise loans

# Culture Jiemian 18m

Chinese media describe alleged fraudsters posing as mobile-company agents who use victims' unlocked phones and face scans to open installment loans for supposedly free handsets.

One 71-year-old Xi'an resident said a caller offered a free phone, then used his unlocked handset and photograph to arrange a 2,155-yuan loan, roughly $300, with monthly payments and interest.
The case illustrates how device subsidies, carrier-style sales scripts, identity verification and consumer credit can combine into a high-friction fraud channel.
The warning is operationally important for companies handling identity workflows: a trusted-brand pretext can turn routine phone access into authorization for financial contracts.
The article presents the residents' accounts and a warning, not a final regulatory finding or adjudicated fraud judgment.
The broader signal is that face recognition and handset control remain socially legible as proof of consent even when older users do not understand the transaction they are authorizing.

Open at Jiemian Ask Claude ChatGPT

China Telecom backs a physical-AI data platform

# AI Leiphone 21h

Maniformer raised several hundred million yuan, roughly $30-60 million, led by China Telecom to scale real-world data collection, governance and evaluation for embodied models.

Its MEgo devices collect environmental and hand-motion data without requiring a robot body, while the company says its pose pipeline reaches millimeter-level trajectory accuracy.
The platform combines collection, spatial reconstruction, multimodal labeling and evaluation, claiming more than tenfold processing improvements and 5% to 10% task-success gains in some feedback loops.
The company has innovation centers in more than 20 Chinese cities and over five overseas nodes, with reported cooperation supplying data to embodied-model projects.
China Telecom's role links cloud, network, physical sites and compliance, suggesting that state-linked infrastructure providers may become distribution channels for embodied-AI datasets.
The reported plan to create tens of millions of hours of interaction data is ambitious, and the supplied text does not establish annotation quality, consent practices or customer revenue.

Open at Leiphone Ask Claude ChatGPT

The White Rabbit revival is part of a broader China nostalgia trend

# Culture SCMP 11h

A viral Western response to White Rabbit candy is driving debate over whether so-called Chinamaxxing reflects genuine interest in China or fatigue with Western culture.

The retro wrapper reportedly drew more than 6 million views and 260,000 likes, turning a mass-market Chinese confection into a piece of design nostalgia.
The trend is less about one candy than about selective cultural import: familiar packaging, everyday products and low-stakes nostalgia can travel more easily than overt political messaging.
The debate also reveals how online audiences construct an idealized China from consumer artifacts, often separating visual and culinary appeal from the country itself.
The article asks whether the movement reflects curiosity about China or a desire to escape Western cultural exhaustion; the supplied extract does not resolve that question.

Open at SCMP Ask Claude ChatGPT

A Chinese container ship sets out on the Northern Sea Route

# Signal Solidot 12h

The Dubai Tower left Ningbo for Britain's Felixstowe via the Bering Strait and Russia's Arctic route, where commercial traffic has grown from 15 ships in 2024 to 23 in 2025.

The voyage is shorter than the Red Sea-Suez alternative when ice conditions permit ships to proceed without an icebreaker.
The increase in ship counts is still small, but the specific voyage shows Chinese operators testing the route as a practical commercial corridor rather than a purely strategic concept.
Environmental groups warn that additional shipping could worsen Arctic ice loss, while warming also makes the route more accessible and more volatile.

Open at Solidot Ask Claude ChatGPT

A former Ecovacs engineer raises $1.5 million for underwater-robotics sensors

# AI 36Kr 45m

Chinese startup Xinyang Micro raised nearly 11 million yuan, roughly $1.5 million, to commercialize CMUT-based sonar modules for pool-cleaning robots and other underwater machines.

CMUT uses MEMS-fabricated vibrating membranes to form dense two-dimensional phased arrays, unlike the larger piezoelectric transducers used in conventional long-range sonar.
The company targets a 0-to-20-meter near-field perception problem, where a robot needs local acoustic imaging for coverage and path planning rather than kilometer-scale detection.
Its first market is pool cleaners because certification and precision demands are lower than in medical ultrasound, allowing the startup to validate both silicon and a complete module sooner.
The reported module combines the CMUT array, an ASIC and software, which is strategically more defensible than selling a bare transducer.
The company says it has entered small-batch production and is working with pool-robot makers, but the reported 95% cleaning coverage is still a company test result.

Open at 36Kr Ask Claude ChatGPT

A Chinese robotics incubator bets on odd consumer machines

# Culture GeekPark 22h

Digua Robotics' closed Demo Day showcased early projects including an AI golf coach, a jumping robot, a knee exoskeleton, a following mini-truck and physical AI companions.

The incubator's thesis is that tangible products can make otherwise hard-to-monetize AI capabilities legible and valuable to consumers, especially in toys, wearables and home devices.
Several teams are choosing narrow engineering wedges: a two-camera golf system, a jump-control module for rough terrain and direct torque prediction for a knee exoskeleton.
The event also shows how Chinese robotics companies are testing overseas markets early, including golf equipment aimed at the US and a modular outdoor vehicle intended for North American users.
The projects are mostly pre-scale, and the article supplies demos, founder backgrounds and aspirations rather than audited sales or reliability data.
As an ecosystem signal, the event suggests China's embodied-AI scene is broadening beyond humanoid platforms into consumer form factors where appearance, privacy and willingness to pay matter as much as model capability.

Open at GeekPark Ask Claude ChatGPT

Chinese startup raises $42 million for digital coating printers

# Signal 36Kr 21h

Shifang Technology raised nearly 300 million yuan, roughly $42 million, to scale high-precision digital printing of functional industrial coatings for batteries, PCBs and cars.

Its UV inkjet engines print functional coatings from one to 120 micrometers thick, with the company claiming up to 99% material utilization and more than 98% first-pass yield in battery-cell projects.
The system combines printheads, control boards, fluid delivery, software and material-process matching, making the moat a full process stack rather than a single nozzle design.
Shifang says its engines have coated more than five million cells and can handle prismatic, cylindrical and blade formats at up to 56 parts per minute.
The battery application replaces or supplements masking and spray processes, while a second growth path targets conformal coatings on PCBs with finer, digitally varied thickness.
The company works with more than 50 materials suppliers, illustrating why industrial printing is as much a materials-integration problem as an equipment problem.
These are company-reported deployment figures, but the combination of local materials supply and high-volume battery manufacturing makes the category strategically relevant.

Open at 36Kr Ask Claude ChatGPT

A simple savings question becomes a debate about how young China should live

# Culture 36Kr 16h

A Burger King promotion turned a second-grade fill-in-the-blank question into a Chinese social-media argument over saving first, traveling now or doing both.

The prompt paired stable saving with traveling the world, and users supplied competing conjunctions that became proxies for delayed gratification, financial anxiety and consumer freedom.
The discussion cuts across age groups: younger workers may have little cash or leave, while older users with more security can afford simultaneous saving and travel.
The debate reflects a broader consumer mood in which money is both a safety buffer and the condition for a life many people feel they are postponing.
The campaign worked because it left the answer open enough for users to project their present financial state onto a trivial language exercise.

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	1/20
Ginger River Review	200	0/20
Sinocism	200	1/20
QbitAI	200	5/10
36Kr	200	8/30
Solidot	200	11/18
Leiphone	200	20/20
GeekPark	200	5/30
IT Home	200	12/60	+29 over cap
SCMP	200	10/50	+6 over cap
Initium	200	4/15
Jiemian	200	8/30	+6 over cap
China Digital Times	200	0/15
The Wire China	200	0/50
Techmeme	200	1/15
Hacker News	200	0/20
V2EX	200	8/50	+34 over cap

Xi took the nineties back this morning; the July data explains why he needed them

Forty minutes for Jiang, then a warning

Unitree lists Wednesday at 219 times earnings

DeepSeek starts charging peak rates today

Video models converge, only one opens the weights

Want Want's sugar problem, and a fill-in-the-blank

Waymo pays 127% and still saves money

Threads we are pulling