China Desk — 24 Aug 2026, 16:50 UTC

Alibaba's liver-cancer model finds overlooked tumors in clinical use

# QbitAI 8h

Alibaba's DAMO LiON model found 15 previously missed malignant lesions in a two-month prospective hospital deployment, suggesting medical AI is moving from retrospective accuracy claims into workflow-level support.

The model reviewed enhanced CT scans for more than 10,000 patients and flagged mostly one-centimeter lesions, many involving easily missed liver metastases.
The study reports that AI-assisted doctors worked 27 percent faster and improved malignant-lesion sensitivity by 11.5 percent, while junior doctors approached senior-doctor performance.
The deployment used disagreement escalation: cases where AI and the initial doctor differed went to senior radiologists and, when needed, a multidisciplinary team.
Technically, the model combines whole-liver context, local texture and multi-phase image fusion to detect low-contrast lesions across difficult liver conditions.
The results were published in Nature Medicine according to the source, but the article does not provide cohort composition, comparator design or longer-term clinical outcomes.
The important business implication is workflow augmentation and missed-diagnosis reduction, not autonomous diagnosis; regulatory and liability boundaries remain decisive.

Open at QbitAI Ask Claude ChatGPT

Xiaomi unveils three chips for phones, edge AI and vehicles

# GeekPark 1h

Xiaomi's three-chip release is less about a phone benchmark than a vertically integrated AI platform: a 3nm phone SoC, a stacked accelerator and a driving chip spanning pocket devices to cars.

Also: IT Home

Xiaomi claims the O3 phone SoC reaches 5.22 million on AnTuTu, uses 24 billion transistors and supports LPDDR6; independent testing is still pending.
The O100 uses wafer-to-wafer stacking of two AI-memory wafers and one NPU wafer at a claimed 1.4 micrometer bond pitch, targeting the memory wall rather than raw compute.
Xiaomi claims 1.22 terabytes per second of O100 bandwidth, about 16 times the cited LPDDR5X bandwidth in a flagship phone, and up to 330 tokens per second in a prototype paired with its MiMo model.
The D100 is presented as a 3nm driving chip with up to 160GB of unified memory and capacity for a local 200-billion-parameter model; an AI Cube prototype runs 120B and 3B models together.
The strategic move is to reuse one compute base across Xiaomi phones, vehicles, appliances and desktop devices, reducing dependence on external chip roadmaps.
The source says Xiaomi has spent 21 billion yuan, roughly $3.0 billion, on chip development over five years; commercialization cost, yield and third-party performance remain unclear.

Xiaomi says chip investment buys its next decade IT Home

Open at GeekPark Ask Claude ChatGPT

Censored essay links banquet assaults to unchecked official power

# China Digital Times 17h

A translated censored essay argues that official sexual-assault cases cannot be addressed case by case because local authorities control when and how the press reports them, exposing a deeper accountability problem.

The essay discusses two alleged assaults by officials at banquets and says reporting on one case disappeared while information about another official was reduced to a removed government biography.
Its central claim is that the problem is not individual impulse but the absence of an institutional cage strong enough to constrain power.
The text describes a familiar Chinese information pattern: outlets wait for an official notice, and reporting without one can be pressured or deleted even when allegations circulate widely.
The essay is a censored blogger's argument, not an adjudication of the underlying allegations, so the value lies in showing how power, censorship and local accountability are understood by critics.
The gap between official disciplinary announcements and independent investigation is itself a governance signal, especially for organizations operating around local officials and state-owned enterprises.
The broader implication is that anti-corruption and workplace-safety campaigns can punish visible offenders without changing the information and incentive structures that permit abuse.

Open at China Digital Times Ask Claude ChatGPT

Xiaomi's humanoid robot returns after two quiet years

# Leiphone 9h

Xiaomi publicly showed its 66-degree-of-freedom humanoid Tianda and said factories, not homes, are the first target; the more important signal is the company's insistence that AI reliability still blocks consumer adoption.

Tianda is about 1.70 meters tall and weighs 66 kilograms, with roughly half of its degrees of freedom in the hands; Xiaomi says the public interaction was model-driven rather than remotely operated.
The robot handled flowers, lightweight incense pieces, handshakes and fist bumps, demonstrations that test fine manipulation but not sustained work.
Xiaomi says it began robotics investment in 2020 and deliberately reduced public announcements after 2023 while continuing internal development.
The factory-first strategy follows a controllable environment, measurable return on investment and easier privacy and safety boundaries; homes are more variable and demand emotional as well as physical competence.
The company's executive says home adoption could take three, five or ten years if the AI does not become reliable, a notably cautious message amid China's robot enthusiasm.
Xiaomi's advantage is internal access to models, manufacturing, supply chain and consumer hardware, but the source gives no production schedule or customer deployment.

Open at Leiphone Ask Claude ChatGPT

Robot executives debate whether brains or factories block scale

# Leiphone 13h

At China's robotics conference, executives split over the bottleneck: some demand 99.9 percent reliability, while delivery veterans say manufacturing consistency, standards and return on investment matter more.

One camp argues that a 96 percent task-success rate is unacceptable in physical work because four failures in 100 can damage a production line or endanger equipment.
Another executive says the company has delivered about 15,000 robots, including more than 10,000 quadrupeds, and that narrow tasks can work with merely adequate intelligence.
Manufacturing leaders point to batch yield, process consistency, mean time between failures and the lack of standardized new-product-introduction procedures as the real barriers to 10,000-unit delivery.
The channel perspective adds a demand problem: robot conversion rates are reportedly far below those of ordinary consumer electronics, suggesting that many products still lack an urgent use case.
The debate reframes embodied AI as a systems-engineering problem spanning model reliability, hardware quality, service cost and customer payback.
The most actionable implication is that a robot company needs a measurable deployment loop and repeatable manufacturing process, not only a stronger demo model.

Open at Leiphone Ask Claude ChatGPT

Shein opens a Hong Kong offering to fund global expansion

# Leiphone 14h

Shein began a Hong Kong share sale targeting about HK$13.12 billion net, or roughly $1.67 billion, with 40 percent earmarked for technology and 40 percent for global branding and expansion.

The offering covers about 279.99 million shares at HK$47.60 to HK$49.50, with trading expected to begin on September 1, according to the filing summary.
The company reports 2025 net revenue of $41.8 billion and net profit of $2.06 billion, serving about 273 million active customers across roughly 160 markets.
Its small-batch automated replenishment model links trend detection, design, supplier allocation, production, routing and customer feedback in a tightly digitized loop.
The listing is strategically important because Shein is using Hong Kong to finance a global supply-chain and brand platform while facing scrutiny over labor, customs, intellectual property and market access.
Shein says its designer and independent-brand accelerator has supported nearly 20 brands, showing an attempt to evolve from a private-label retailer into a broader marketplace.
The prospectus numbers describe management's claims and plans; the market will test whether speed and low prices can coexist with higher compliance and brand-building costs.

Open at Leiphone Ask Claude ChatGPT

A Chinese humanoid robot plays autonomous tennis on live television

# GeekPark 9h

Galaxy General's robot returned serves, moved across the court, recovered from falls and coordinated in doubles, offering a more demanding physical-AI demonstration than choreographed walking or dancing.

The match required perception, positioning, stroke selection, balance and recovery under changing ball speed and trajectory rather than a fixed action script.
Galaxy General says its model combines high-level game understanding, lower-level motion control and neural control in one architecture.
Its training platform converts imperfect human data into usable trajectories and uses virtual multi-agent tennis matches to generate large numbers of physical interactions.
The reported system is claimed to transfer from simulation to the physical court, which is strategically important because high-quality teleoperation data for tennis is scarce.
Live competition remains a narrow domain; the source does not disclose latency, failure rates, human intervention or performance across different courts and opponents.
The demonstration matters because it tests closed-loop action and recovery, not simply visual realism or a pre-recorded performance.

Open at GeekPark Ask Claude ChatGPT

Finance veteran Li Yun takes Shenzhen's top city posts

# SCMP 7h

A finance veteran became Shenzhen's deputy Party secretary and acting mayor ahead of Apec meetings, signaling a preference for financial-management experience at a strategically exposed city.

Li Yun spent more than two decades in financial institutions and had served as a Guangdong vice-governor since March 2025, according to the report.
He replaces Qin Weizhong in the acting-mayor role and assumes the deputy Party-secretary position, linking municipal administration and Party management.
The appointment comes before Shenzhen hosts a series of Apec meetings, when the city will be expected to project economic openness and policy competence.
Shenzhen's leadership changes are worth tracking because the city sits at the intersection of technology, finance, Hong Kong integration and national industrial policy.
The source does not identify a policy program, so the appointment should be read as a personnel signal rather than proof of an imminent economic shift.

Open at SCMP Ask Claude ChatGPT

Former education founder raises nearly $1.4 billion for home robots

# 36Kr 15h

After China's education crackdown destroyed his first company, Zhang Yi spent four years building an affordable home robot with modular models, world-model prediction and real household data.

The source says his former education company once employed about 100,000 people before the policy shift forced mass layoffs and eventual delisting, an unusually direct example of regulatory risk reshaping an entrepreneur's direction.
Future Is Near reportedly raised nearly 1 billion yuan, roughly $141 million, across three rounds in 2026 after Zhang self-funded more than 100 million yuan, roughly $14 million.
The team iterated 16 machines, targeted a consumer price of 20,000 to 30,000 yuan, roughly $2,800 to $4,200, and offered household trials at about 3,000 yuan, roughly $423, for rental.
Its AVLA architecture decomposes long tasks into specialist models and uses a dispatcher, while a world model predicts candidate action paths before execution.
The company says its Self-Evolving World Action Model feeds household outcomes and failures back into training, using corner cases that factories cannot reproduce.
About 500 Shanghai households are reportedly using or testing the robot, but the source provides no independent success rate, service cost or repeat-purchase data.
The larger lesson is that home robotics may be a data strategy as much as a product strategy, but capital raised before reliable unit economics can become a liability when the market cools.

Open at 36Kr Ask Claude ChatGPT

ByteDance consolidates its AI work tools under Doubao

# 36Kr 8h

Chinese coverage stresses a new Doubao Work brand and Feishu integration, while US coverage simply describes consolidation; the move makes productivity agents a core ByteDance bet.

Also: Techmeme

Trae Work and Coze capabilities are being folded into Doubao, while Trae's IDE and CLI continue as coding products under the Doubao brand.
The move follows ByteDance's integration of Feishu and Doubao and recent additions such as remote computer control, virtual desktops and more than 200 connectors and skills.
Centralizing these products can improve distribution and shared model infrastructure, but it also risks losing specialist identities and developer trust built by Trae and Coze.
ByteDance's broader strategy reportedly prioritizes AI, information platforms and transaction services, making Doubao a designated main business rather than an experimental app.
The source includes a company response saying existing user rights will remain unaffected, but it does not specify final product architecture or pricing.

ByteDance folds Trae and Coze into Doubao Techmeme

Open at 36Kr Ask Claude ChatGPT

A new benchmark tries to measure AI that can do real science

# AI QbitAI 3h

A Chinese AI-for-science company helped propose an evaluation framework that scores the full research loop, replacing answer-only exams with evidence of hypothesis, experiment and interpretation.

The proposed discovery-episode framework evaluates hypothesis formation, experimental execution, result interpretation and the end-to-end discovery process.
It treats failed experiments as training and evaluation data, because a useful scientific system must recognize anomalies and update its plan rather than produce plausible prose.
The framework was developed with Microsoft Research, Stanford and other institutions, according to the source, giving it broader participation than a single-company benchmark.
Deep Principle's MIRA platform is presented as an industrial example linking hypothesis generation, simulation, wet-lab validation and knowledge retention.
The company claims first place on two named benchmarks and reports that its React-OT model computes chemical reaction transition states in 0.4 seconds; the article does not provide independent validation or benchmark protocols.
The strategic question is whether AI-for-science companies can own the closed-loop experimental data that becomes more valuable than generic language-model output.

Open at QbitAI Ask Claude ChatGPT

HiDream presents a native all-modal world model

# AI QbitAI 11h

HiDream's O1 World model combines text, images and interactive control, aiming to turn generated scenes into persistent worlds that can be explored, edited and used for robot simulation.

Also: Leiphone

The company uses a Unified Transformer architecture to combine modalities and claims top placement in a WBench sub-ranking on its first evaluation.
Its stated capabilities are roaming, editing and interaction, with long-horizon spatial and physical consistency as the main technical target.
The proposed data flywheel links internet video, simulated physical trials and real interaction feedback to world models, agents and robot bodies.
The approach is strategically different from a video generator because the model must retain scene state and respond to actions over time.
The source is a company presentation and offers no independent evaluation of memory, physical correctness or intervention rates.

HiDream's world model aims to connect simulation with action Leiphone

Open at QbitAI Ask Claude ChatGPT

Chinese AI-chip maker Enflame starts its Star Market offering

# AI Leiphone 4h

Enflame has begun the Shanghai Star Market listing process after building four chip architectures and a full software-to-cluster stack, making commercialization rather than tape-out the central test.

The company reported revenue of 301 million yuan in 2023, 722 million yuan in 2024 and 990 million yuan in 2025, approximately $42 million, $102 million and $139 million.
It reported 1.12 billion yuan, roughly $158 million, in first-half 2026 revenue and expects 2026 full-year or 2027 consolidated profitability, according to the prospectus summary.
Enflame claims 3.676 billion yuan, roughly $518 million, of research spending from 2023 through 2025, equal to more than its cumulative revenue over that period.
Its GCU architecture, chip-to-chip interconnect and TopsRider software target the migration problem that often blocks domestic accelerators after silicon works.
The proposed 6 billion yuan fundraising, roughly $845 million, is aimed at fifth- and sixth-generation chips and hardware-software co-design.
The filing is valuable because it exposes revenue, spending and commercialization claims, but the company still faces ecosystem and customer-concentration risks against entrenched foreign platforms.

Open at Leiphone Ask Claude ChatGPT

Alibaba's Wan 3.0 video model launches with document input

# AI QbitAI 11h

Wan 3.0 can generate up to 30-second clips and accept office documents as references, while Chinese users and companies emphasize stable continuity over spectacular single frames.

Also: Leiphone

The model supports document formats including DOC, XLS, PPT, PDF and Markdown, extending video generation from prompts and images toward production briefs and source materials.
Alibaba claims more stable characters, objects, voices, spatial relationships and style across a 30-second sequence, the qualities that determine whether a clip can enter a real workflow.
The model is already being used in short dramas, advertising, tourism and music-video production according to the company, though the article supplies no customer names or output volumes.
API pricing is temporarily discounted to 0.21, 0.42 and 0.84 yuan per second for 480p, 720p and 1080p, roughly 3, 6 and 12 US cents per second.
The significant product direction is controlled, repeatable production rather than novelty generation; independent comparisons and rights-management details are missing.

Wan 3.0 video model enters Alibaba's production stack Leiphone

Open at QbitAI Ask Claude ChatGPT

PDD presents its second quarter as a governance-led reset

# Signal Leiphone 6h

Chinese coverage stresses PDD's quality-growth program and platform governance, while US coverage highlights an $16.72 billion revenue miss and a 12 percent profit decline.

Also: Techmeme

PDD reported second-quarter revenue of 112.4 billion yuan, roughly $15.8 billion, and profit of 27.2 billion yuan, roughly $3.8 billion, in the Chinese account.
Non-GAAP research spending reached 4.3 billion yuan, roughly $606 million, up 40 percent year over year.
Management says more than 150 governance measures now cover product admission, food and drug safety, advertising, intellectual property, livestreaming and algorithm abuse.
The company also describes a 4,000-person operation in Xiongan focused on data processing and traditional-manufacturing services, linking e-commerce to industrial upgrading.
The positive domestic framing sits alongside the US account's weaker growth and profit data, making competition from livestreaming and social commerce the key unresolved question.

PDD revenue misses as competition pressures Temu owner Techmeme

Open at Leiphone Ask Claude ChatGPT

Domestic inference chip maker HL200 targets 10,000-GPU clusters

# AI QbitAI 9h

Zhongcheng Hualong launched an inference chip and cluster design spanning eight cards to 10,240 cards, but the company's performance claims are vendor-reported and its enterprise deployments need scrutiny.

HL200 claims FP4, FP8 and FP16 support, with 4 petaFLOPS at FP4, 2 petaFLOPS at FP8 and 0.5 petaFLOPS at FP16 or BF16.
The architecture is paired with PyTorch, ONNX and vLLM compatibility plus an OpenAI-compatible interface, targeting migration cost rather than only peak silicon performance.
The proposed topology supports 64 GPUs per rack, up to 1,024 cards in a node and up to 10,240 cards horizontally, with compute, storage and networking treated as one service-level system.
The company says benchmark adaptations for several Chinese large models show prefill and decode advantages over foreign equivalents, but the source supplies no reproducible workloads or independent comparisons.
The announcement reflects China's shift from single-chip nationalism to full-stack procurement stories, where software, cluster management and delivery matter as much as chip design.
The source is sponsored content and claims deployments with major telecom and infrastructure customers, so commercial validation remains the key missing evidence.

Open at QbitAI Ask Claude ChatGPT

Chinese government vessels near Taiwan set a third straight monthly record

# Policy SCMP 1h

Taiwan recorded 118 mainland government-vessel sightings in the first 24 days of August, exceeding the previous two months; the pattern points to sustained law-enforcement pressure rather than a one-off exercise.

Taiwan's defense ministry reported 118 sightings in August through day 24, compared with 117 in July and 111 in June.
The vessels include coastguard and survey ships, indicating a widening official presence around Taiwan rather than only military signaling.
The report links the increase to Beijing's response to maritime boundary discussions involving Japan and the Philippines, but that causal interpretation is analytical rather than an official explanation.
Repeated coastguard activity can normalize contested jurisdiction claims while keeping escalation below the threshold of overt military operations.
For operators tracking regional risk, the important variable is persistence: a third consecutive record suggests an institutionalized operating pattern.

Open at SCMP Ask Claude ChatGPT

DeepSeek adds a multimodal API with unusually low image-token usage

# AI GeekPark 9h

DeepSeek's experimental vision API adds image understanding and file reuse while charging the same token rates as its text model, making multimodal access cheaper to prototype even before quality is independently validated.

The API accepts base64 images, URLs and uploaded files, while Files API identifiers allow repeated calls without retransmitting the same image.
DeepSeek claims a maximum of 384 tokens per image, below the 800 to 1,100 tokens cited for comparable processing by GPT and Claude systems.
The source attributes the efficiency to a visual-primitives approach that represents image elements spatially rather than expanding every image into dense patches.
The model is described as matching the text-only model on agent, reasoning and world-knowledge benchmarks, while its multimodal agent performance is claimed to approach a leading competitor.
A hands-on test generated an HTML game from a screenshot in 36 seconds, but that is a demonstration of code generation rather than evidence of production reliability.
The strategic implication is that low-cost visual input can make image-aware agents economically viable in China, where developers are highly price sensitive.

Open at GeekPark Ask Claude ChatGPT

51Sim takes its driving-simulation stack into robots and space

# Analysis Leiphone 8h

Chinese simulation company 51World is extending a decade of autonomous-driving digital-twin work into robotics, low-altitude flight and space, while admitting the latter markets are not yet commercial.

Its autonomous-driving business generated 51.238 million yuan, roughly $7.2 million, in first-half 2026 revenue, up 545.2 percent year over year and 41.4 percent of company revenue.
The new robot products separate data capture from simulation: AperData collects and processes robot trajectories, while AperOne reconstructs environments for training, testing and evaluation.
The company claims trajectory physical consistency of 99 percent and more than tenfold higher collection efficiency than traditional teleoperation, but the source provides no independent test protocol.
Its neutral position as a simulation and validation supplier lets it serve multiple robot makers without competing directly in robot bodies or data sales.
The expansion illustrates a credible platform thesis, but the space and deep-space projects remain exploratory and lack a demonstrated revenue model.

Open at Leiphone Ask Claude ChatGPT

China and Jordan call for normal shipping through Hormuz

# Policy SCMP 3h

Beijing and Amman jointly called for a ceasefire and restored shipping through Hormuz, using a royal visit to frame China as a diplomatic advocate for sovereignty and negotiated regional order.

The joint statement calls for reopening the strait and maintaining the US-Iran ceasefire while addressing the conflict's root causes through dialogue.
Its language emphasizes sovereignty, security and territorial integrity, themes that align with China's standard diplomatic vocabulary and its opposition to externally imposed regional settlements.
The statement does not show that China can resolve the crisis, but it demonstrates how Beijing uses bilateral diplomacy to position itself as a stabilizing alternative to US-led security management.
Shipping, energy and food-supply disruption give the statement commercial relevance beyond Middle East diplomacy.

Open at SCMP Ask Claude ChatGPT

TikTok and ByteDance agree to a $400 million child-privacy settlement

# Policy Solidot 7h

TikTok and ByteDance will pay $400 million to resolve US allegations that children under 13 could create accounts and have data collected without parental consent.

The settlement follows a 2024 US Justice Department and Federal Trade Commission lawsuit alleging violations of the Children's Online Privacy Protection Act.
The reported payment is split into $300 million immediately and $100 million after a prior consent order involving Musical.ly is withdrawn.
The case shows that TikTok's legal exposure is not limited to national-security review; child privacy and platform-design practices remain major liabilities for ByteDance's US business.
A settlement resolves the litigation without establishing every allegation at trial, but its size makes children's data governance a material operating cost.

Open at Solidot Ask Claude ChatGPT

Chinese team develops a compact real-time X-ray ghost camera

# Signal SCMP 12h

Researchers in China demonstrated tabletop X-ray ghost imaging for high-speed see-through video with potentially lower radiation and simpler hardware than conventional methods.

The technique reconstructs moving objects from correlated measurements rather than forming a conventional frame directly, allowing it to track dynamic structures.
The reported applications include rotating aircraft engines and biological motion such as a beating heart, where motion blur and radiation exposure complicate ordinary imaging.
The source does not provide resolution, frame rate, dose, reconstruction latency or a comparison against clinical or industrial systems.
If the claims survive replication, the architectural advantage would be a smaller sensing setup for dynamic imaging rather than merely sharper still images.

Open at SCMP Ask Claude ChatGPT

Former security guard reaches the AI contest finals

# Culture QbitAI 17h

A former security guard built a browser-based neck-exercise game with TRAE, while the contest's hardware winner linked natural-language design to PCB fabrication; the stories show AI's reach beyond programmers.

Also: 36Kr

The developer says he went from short-video effects work to building the game alone in two months without English fluency or formal software training.
The winning hardware project generated component choices, pin maps, code, simulation and PCB machining, then revised the system after testing it with about 500 real users.
The important evidence is not the prize money but the feedback loop: missing component data, broken builds and hardware-software mismatches forced the team to make AI outputs manufacturable.
The account is closely tied to TRAE promotion, so it demonstrates a plausible workflow rather than proving that nontechnical users can reliably ship arbitrary hardware.

AI coding contest puts ordinary Chinese creators on stage 36Kr

Open at QbitAI Ask Claude ChatGPT

ByteDance explores autonomous driving under its world-model team

# AI 36Kr 58m

Chinese reporting says ByteDance is preparing an autonomous-driving effort around its Seed world-model group, while ByteDance denies plans for an intelligent-driving business.

The reported target is unmanned logistics, potentially connected to ByteDance's automotive cloud business, rather than a consumer passenger-car program.
The proposed team would sit near Seed's multimodal, world-model and robotics research, creating a path from general physical AI into vehicle control.
ByteDance's denial matters: this could be an early research program, talent recruitment effort or internal exploration rather than a committed product line.
The technical rationale is data reuse: autonomous driving offers large amounts of physical-world video and action data that could improve world models and later robotics systems.
A serious entry would bring ByteDance substantial compute, recruiting power and cloud distribution, but the source offers no vehicle prototype, regulatory filing or confirmed customer.

Open at 36Kr Ask Claude ChatGPT

China's view of Trump shifts from alarm toward resignation

# Analysis SCMP 4h

A Chinese analyst argues that official and scholarly views of Trump's second term have moved from alarm to a shrug, reflecting confidence that US volatility can be managed rather than reversed.

The analysis says Chinese observers increasingly see Trump as weakening US alliances and making American policy less predictable.
That does not mean Beijing views the United States as harmless; it suggests a shift from expecting stable competition to planning around repeated shocks.
The piece is explicitly an account of officials and scholars rather than a survey of all Chinese citizens, so it should not be read as public-opinion measurement.
The framing reveals a common Chinese policy instinct: convert foreign political volatility into an argument for domestic resilience and longer planning horizons.

Open at SCMP Ask Claude ChatGPT

Taiwan seeks $1.13 billion for its Strong Bow missile

# Policy SCMP 4h

Taiwan disclosed details of its locally developed Strong Bow interceptor while seeking $1.13 billion for mass production, signaling a move from prototype capability toward layered air defense.

The system reportedly uses hypersonic flight, two-stage solid-fuel propulsion and active guidance, although the source provides only a partial specification set.
The funding request is tied to production, not merely research, making industrial capacity and replenishment as important as the missile's performance.
The program is an adjacent China story because the stated rationale is growing pressure from Beijing, while the immediate actor is Taiwan's defense establishment.
Disclosure of specifications can support deterrence but also gives Beijing more information about the island's defensive architecture.

Open at SCMP Ask Claude ChatGPT

Why the rough Chinese film Bull Came became a hit

# Culture Initium 18h

A Chinese cultural analysis asks why the deliberately rough film Bull Came spread online, treating its imperfections as participatory entertainment rather than production failure.

The article frames the film's appeal around viewers turning visible amateurism, shared jokes and low-budget craft into a collective internet event.
Its significance is cultural: audiences can reward sincerity, absurdity and remixability even when a work lacks conventional polish, a pattern relevant to short-video and AI-generated media.
The source is an essay rather than box-office reporting, so it explains a reception theory rather than proving why every viewer watched the film.

Open at Initium Ask Claude ChatGPT

US and Chinese diplomats trade espionage accusations in Latin America

# Policy SCMP 3h

US and Chinese diplomats are publicly arguing over Chinese technology firms in Latin America, showing that infrastructure-security competition is moving deeper into the Western Hemisphere.

The US official named Huawei, ZTE, Hytera, Hikvision, Dahua and DJI when warning of alleged espionage risks, according to the report.
Chinese diplomats responded on social media, turning vendor security claims into a broader contest over sovereignty, development and US influence.
The exchange matters because Latin American telecom, surveillance and drone procurement can create long-lived technical dependencies and data-access relationships.
The report describes rhetoric and allegations rather than a new ban, procurement decision or verified espionage incident.
For vendors, the strategic risk is that ordinary commercial deployments become bundled into national-security narratives and diplomatic retaliation.

Open at SCMP Ask Claude ChatGPT

China delays the Chang'e 7 lunar mission to next year

# Signal Solidot 7h

China postponed Chang'e 7 because launch conditions were not judged ready, with no official reason given; the mission's search for lunar polar water ice now moves to a later window.

The official statement invokes reliability and a zero-failure standard but does not identify a technical fault.
The planned mission targets the lunar south pole, where lighting conditions matter for landing and surface operations.
The likely next launch window is February or March 2027, though that timing is not confirmed by the official notice.
The delay shows China's space program prioritizes schedule flexibility over publicly explaining setbacks, limiting outside assessment of program health.

Open at Solidot Ask Claude ChatGPT

Humanoid robots outrun humans but struggle to stop

# Signal Solidot 1d

China's humanoid games produced robot sprint times faster than human records, yet crashes at the finish show why braking, recovery and reliability matter more than peak speed.

Also: IT Home

Two robots reportedly beat the 9.58-second human 100-meter record, and another ran 400 meters in 39.7 seconds, but several hit safety pads because braking remained poor.
More than 40 percent of the event's contests require full autonomy, while other tests cover packaging, cable connection, offices, restaurants and emergency scenarios.
The contrast exposes a classic physical-AI failure mode: optimizing a narrow objective such as speed without the full control envelope needed for safe deployment.
For industrial buyers, stopping distance, damage tolerance, recovery after falls and repeatable task completion are more useful metrics than a headline race time.
The event is valuable precisely because its public failures reveal engineering gaps that polished demonstrations hide.

Humanoid robot wins China's first competition weightlifting title IT Home

Open at Solidot Ask Claude ChatGPT

Robot startup founder says automakers are too early for embodied AI

# Analysis 36Kr 31m

Chinese coverage stresses the gap between embodied-AI funding and mass production, while the founder argues automakers will enter eventually but are mostly making tactical bets today.

Kunlunxing was founded by former Huawei and Alibaba executive Reng Geng and former Li Auto driver-assistance chief Lang Xianpeng; the source says it raised tens of billions of yuan, or several billion dollars, across three rounds in under 90 days.
The company is pursuing a full-stack humanoid approach modeled on Tesla Optimus, with the stated goal of mass production rather than demonstrations.
Reng divides embodied intelligence into humanoid, humanoid-like and non-humanoid forms, arguing that shared AI and hardware capabilities can eventually support all three.
His criticism of automakers is organizational as much as technical: a small internal department lacks startup-level ownership, while corporate controls over computing, hiring, legal work and finance slow iteration.
The strategic test is whether the company can turn a reusable stack into reliable commercial deployments; the article provides ambition and fundraising claims but no production volume or customer evidence.

Open at 36Kr Ask Claude ChatGPT

China's Science Exploration Award adds its first post-1995 winner

# Culture Leiphone 3h

China's major young-scientist award selected 50 researchers, including its first post-1995 winner and 10 women; the changes show an effort to widen and professionalize elite research funding.

Each winner receives 3 million yuan over five years, roughly $423,000, with unrestricted use, according to the source.
The award says 10 women won this year, while international reviewers now make up more than 40 percent of peer review, suggesting an attempt to broaden both recipients and evaluators.
The program introduced a mandatory gap year after three unsuccessful applications, a governance change intended to reduce repetitive grant-chasing.
Joshua Zahl, a Canadian mathematician working at Nankai University, is highlighted as evidence that stronger domestic research networks and funding can attract foreign researchers.
The award is backed by a 10 billion yuan foundation, roughly $1.4 billion, funded by Tencent, and has supported 397 researchers over eight years.

Open at Leiphone Ask Claude ChatGPT

China turns coal into higher-value liquid fuels

# Signal SCMP 6h

A Ningxia plant has begun full operation converting coal into oil products, illustrating China's willingness to industrialize carbon-intensive routes when they serve energy and strategic-material goals.

The plant processed 24 million tonnes of coal in 2025, about a quarter of Ningxia's annual output, according to the cited Chinese report.
The resulting products can support rocket fuel and industrial lubricants, allowing coal to enter higher-value chemical and aerospace supply chains rather than only power generation.
The reported sevenfold increase in economic value is a source claim, not an independently audited margin calculation.
The project shows how China's energy-security logic can override simple decarbonization narratives by combining domestic coal abundance with chemical substitution.

Open at SCMP Ask Claude ChatGPT

Space thermal-management startup raises an A-plus round

# Signal 36Kr 6h

Beijing thermal-control supplier Ruilai raised tens of millions of yuan, roughly several million dollars, as high-power computing satellites make active liquid cooling a core spacecraft constraint.

The company says its liquid-cooling systems hold 90 percent of the domestic market and that one satellite system has operated in orbit for more than three years.
Its stack spans pumps, fluid loops, radiators, heat pipes, vapor chambers and coatings, allowing it to control interfaces that are often split across suppliers.
The source contrasts passive radiators and heat pipes with pumped liquid loops needed when satellite power reaches tens or hundreds of kilowatts.
The company expects five high-power computing satellites using its systems by the end of 2026, but that is a forecast rather than delivered fleet scale.
The strategic bottleneck is validation: an orbital cooling loop cannot be repaired, so years of telemetry and reliability modeling matter more than a ground demonstration.
Ruilai's attempt to combine aerospace-grade components with commercial manufacturing could reduce cost, but the source offers no audited failure-rate or unit-economics data.

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	0/20
QbitAI	200	8/10
36Kr	200	10/30
Solidot	200	14/20
Leiphone	200	19/20
GeekPark	200	9/30
IT Home	200	12/60	+42 over cap
SCMP	200	10/50	+2 over cap
Initium	200	3/15
Jiemian	200	8/30	+14 over cap
China Digital Times	200	1/15
The Wire China	200	0/50
Techmeme	200	4/15
Hacker News	200	0/20
V2EX	200	8/40	+25 over cap

Hong Kong funded one China today and marked the other one down

Alibaba asks shareholders for the compute bill

Shein is worth $27bn, and the discount is the tariff

Xiaomi's three chips meet a memory squeeze

Shenzhen gets a banker, TikTok pays $400m

A master's degree or the assembly line

A censored post names the mechanism, not the man

Threads we are pulling