China Desk — 15 Aug 2026, 12:54 UTC

Alibaba open-sources Qwen3.8-27B for local deployment

# QbitAI 21h

Chinese coverage stresses Qwen3.8-27B’s Apache 2.0 release, 262K native context and consumer-GPU deployment, while US forum discussion focuses on practical local coding and benchmark caveats.

Also: Leiphone, Hacker News

Alibaba says the dense multimodal 27-billion-parameter model improves coding and office work, adds a reasoning-effort control and can extend its 262K context to 1 million tokens with YaRN.
The Apache 2.0 weights target self-hosting and commercial use, making the model strategically relevant to teams that want local inference and control over data.
Chinese coverage emphasizes the breadth of Alibaba’s open-model ecosystem, including more than 460 released models, over 3 billion global downloads and more than 300,000 derivative models; these are company-reported figures.
In the linked Hacker News discussion, developers describe useful local coding assistance on Apple and Nvidia hardware but also report low throughput, regressions on some benchmarks and reservations about headline comparisons.
The real engineering question is the cost-quality frontier: enough capability on a local GPU can be more valuable than a stronger hosted model when privacy, latency or predictable spend matter.
This release matters because it is both a model event and a distribution strategy: open weights let Alibaba compete for developer workflows beyond its own cloud.

Alibaba says Qwen3.8-27B is ready for consumer GPUs Leiphone
Hacker News tests Qwen3.8-27B as a local coding agent Hacker News
Qwen3.8-27B reportedly runs on a consumer GPU QbitAI

Open at QbitAI Ask Claude ChatGPT

Tencent’s AI spending cut quarterly profit growth to 9%

# GeekPark 4h

Tencent’s second-quarter AI buildout had a reported 10.5 billion yuan net impact on profit (roughly $1.5 billion), while capital spending reached 52.8 billion yuan (roughly $7.4 billion), turning AI from a promise into a financial risk.

Tencent reported 204.79 billion yuan in quarterly revenue (roughly $28.8 billion) and 75.64 billion yuan in non-IFRS operating profit (roughly $10.7 billion), but says new AI products reduced operating profit by about 10.5 billion yuan.
Capital expenditure rose 176 percent year over year to 52.78 billion yuan, and research spending rose 35 percent to 27.28 billion yuan (roughly $3.8 billion); free cash flow turned negative 13.8 billion yuan (roughly negative $1.9 billion).
Management says most new compute is being reserved for internal models and applications before any excess is rented through Tencent Cloud, prioritizing strategic control over immediate utilization revenue.
WorkBuddy, Tencent Cloud and token sales are presented as the nearest commercial links, while the WeChat AI assistant is a larger option on agent-mediated transactions and service discovery.
The company argues that compute is a fallback asset that could be rented profitably if internal products disappoint, but that does not eliminate opportunity cost or depreciation risk.
The key question is whether first-party models and distribution create more incremental value than selling the same compute capacity to outside customers.
For a Chinese platform company, the spending also signals that AI is being treated as infrastructure for future product surfaces, not just as a feature added to existing apps.

Open at GeekPark Ask Claude ChatGPT

GLM-5.3 targets coding and security through post-training

# Leiphone 1d

Chinese lab Zhipu says GLM-5.3 keeps the same base model but sharply improves coding and vulnerability discovery through post-training, including a reported 2,436 flaws across 269 projects.

Zhipu attributes the gains to post-training methods and frameworks including IndexShare, the SAO reinforcement-learning algorithm and its Slime training system, rather than a new base architecture.
The company reports a Terminal-Bench 3.0 score rising from 4.6 to 28.3 and a DeepSWE score rising from 46.2 to 66.9, while claiming a lead on several agent and workplace evaluations.
Its internal coding evaluation places the model in realistic development environments with code, documents, compute and storage, testing sustained tool use rather than isolated code completion.
Zhipu says joint security work found 2,436 vulnerabilities in 269 open-source and commercial projects, including 1,097 medium- or high-severity flaws; these are vendor-reported results without disclosed methodology in the item.
The safety design uses layered filtering, an intent-aware reasoning monitor and model-level alignment, a particularly important choice for downloadable weights where hosted-service controls disappear.
The release suggests Chinese labs are shifting competition from parameter count to post-training, tool use and secure agent behavior.
The open-model claims should be checked against reproducible benchmarks and real repository fixes before treating them as evidence of a durable lead.

Open at Leiphone Ask Claude ChatGPT

Beijing issues ecological reports to reinforce South China Sea claims

# SCMP 1h

Chinese authorities published ecological assessments for nine disputed South China Sea features, using environmental monitoring to reinforce a territorial position amid a dispute with Washington.

The Ministry of Natural Resources’ South China Sea bureau released a 2025 marine early-warning bulletin alongside a protection and restoration report for the disputed region.
Environmental reporting is not politically neutral here: defining monitoring, restoration and protected areas helps Beijing present administrative control as stewardship.
The move follows an escalating dispute over China’s establishment of a national nature reserve in contested waters, linking ecological governance to sovereignty claims.
The reports give China a bureaucratic and scientific vocabulary for its position, potentially complicating foreign challenges that frame the issue only as militarization.
Their practical importance depends on whether the claimed reserve changes access, enforcement or activity by other coastal states.

Open at SCMP Ask Claude ChatGPT

Beijing worries about Washington chaos before Xi’s planned visit

# SCMP 13h

Chinese officials are reportedly unsettled by disorder in the Trump administration less than six weeks before Xi Jinping’s planned Washington visit, reviving memories of earlier scheduling failures.

The report says poor coordination frustrated Chinese officials before an earlier Trump visit and that Beijing now sees similar uncertainty around Xi’s first planned state visit to Washington since 2015.
For Chinese diplomacy, logistics are political: uncertainty about the host’s agenda and authority can signal whether commitments made at the summit will survive internal US competition.
The concern is not simply protocol; it affects preparation, messaging and the ability to trade concessions across security, technology and economic issues.
The item relies on unnamed sources and describes a planned visit, so both the timing and the level of official anxiety could change.
The strategic signal is that Beijing may treat US domestic disorder as a negotiating variable rather than merely an American internal problem.

Open at SCMP Ask Claude ChatGPT

Zhu Rongji’s death reopens debate over China’s lost reform era

# SCMP 23m

Chinese coverage mourns Zhu Rongji as a bridge to the United States, while another Chinese analysis stresses that his reforms also strengthened state control.

Also: The Wire China

The source presents Zhu as a premier willing to take personal risks for better US-China relations and as a global figure before China’s current power status.
His appeal abroad was tied to unusually direct communication and a belief that productive bilateral relations served both countries.
The contrast within Chinese-origin coverage matters: one frame emphasizes diplomatic openness, while another treats administrative reform as a mechanism for centralizing power.
That tension captures why Zhu’s legacy remains useful for understanding China’s governing model: market reform and stronger state capacity were not opposites in his era.
The story matters most as a lens on what Chinese and foreign audiences remember from the period before strategic rivalry hardened.

Zhu Rongji’s reforms strengthened the Chinese state The Wire China

Open at SCMP Ask Claude ChatGPT

Chinese quantum startup raises more than 100 million yuan

# 36Kr 11h

A University of Science and Technology of China spinout raised more than 100 million yuan (roughly $14 million) to develop integrated photonic quantum computers, specialized optical compute and quantum-randomness chips.

Silicon Zhen says it has built a cloud-accessible programmable photonic quantum computer and internally validated a 5,000-bit specialized optical-computing device.
The company reports generating a four-photon, 16-qubit GHZ state on chip and running Grover’s algorithm with a 98.7 percent average identification probability; these are company-linked research claims, not independent product benchmarks.
Its near-term commercial product is a quantum random-number generator chip, with about 200,000 units ordered and use cases including smart cards, utility meters and home cameras.
The company’s strategy is staged: quantum-randomness chips provide revenue, specialized optical compute is expected to follow, and a million-qubit universal machine remains a long-term target.
Photonic systems can operate at room temperature and use silicon manufacturing, but scaling still depends on optical loss, fabrication tolerances, interconnects and deep foundry collaboration.
The funding is strategically relevant because it combines university-originated research, industrial capital and a plan to turn a speculative computing path into nearer-term security hardware.

Open at 36Kr Ask Claude ChatGPT

ByteDance chooses slow model-building over distillation

# Jiemian 28m

ByteDance is accepting a short-term AI gap while building its own stack, creating a new data-and-safety division, considering a model above 5 trillion parameters and accelerating product monetization.

ByteDance founder Zhang Yiming and CEO Liang Rubo reportedly told staff that Seed would not use distillation as a shortcut, even if its models temporarily lag domestic rivals.
The company created an AI Data and Safety first-level department with more than 1,000 data workers, parallel to Seed, Flow and Doubao, making data governance a peer to model and product organizations.
A model larger than 5 trillion parameters is reportedly under discussion, but the item does not establish that training has begun or that scale alone would improve capability.
Doubao is adding paid subscription tiers, transaction commissions in life services and a closer relationship with Feishu, pointing toward an integrated consumer, enterprise and agent business loop.
The strategic bet is expensive and slow: proprietary data pipelines and post-training may create durable capabilities, but only if ByteDance can convert them into distribution and revenue before rivals close the gap.
The story is valuable because it reveals a Chinese lab choosing organizational and data infrastructure over benchmark imitation, rather than simply announcing another model release.

Open at Jiemian Ask Claude ChatGPT

A weight-decomposition method aims to make model circuits cheap to inspect

# AI QbitAI 6h

Chinese researchers propose Sparse Weight Decomposition, extracting sparse, independently ablatable units directly from trained weights and using less than 1% of the calibration data of trained surrogate methods.

SWD approximates a dense weight matrix as two sparse matrices, with shared intermediate coordinates acting as bottleneck units that can be scored, selected and removed individually.
Unlike a dense singular-value decomposition, each SWD unit has sparse read and write connections, making a selected circuit more interpretable as an actual computational path rather than merely a low-rank component.
The method optimizes output reconstruction on a small calibration set and alternates factor updates while enforcing a hard nonzero budget.
The paper reports tests on GPT-2 Small and several Qwen models, where SWD reached comparable replacement fidelity with fewer than 1 percent of the tokens used by the cited training-based baselines.
That result addresses a practical bottleneck in mechanistic interpretability: scaling analysis across layers and checkpoints without training a new explanatory network for every target.
The important caveat is that faithful matrix reconstruction is not the same as proving a semantic circuit; sufficiency, necessity and causal intervention still determine whether the extracted units explain behavior.
If the reported results generalize, direct weight factorization could make large-scale circuit analysis much cheaper and more reproducible.

Open at QbitAI Ask Claude ChatGPT

Two Chinese mathematicians win Fields Medals, sparking a debate at home

# Culture China Digital Times 13h

Two Chinese mathematicians won Fields Medals for the first time as a pair, prompting pride and an online argument over elite training, overseas careers and support for young researchers.

Wang Hong, the first Chinese woman to win the medal, was recognized for work in harmonic analysis and geometric measure theory, including the Kakeya conjecture.
Deng Yu was recognized for partial differential equations and work related to Hilbert’s sixth problem; both developed their careers through universities in the United States and Europe.
Chinese social media celebrated the achievement while debating how much credit belongs to Peking University and whether domestic institutions give emerging mathematicians enough encouragement.
The dual success exposes a familiar tension in Chinese science: national prestige is claimed for globally trained talent even when the strongest research careers remain institutionally international.
The debate is more revealing than the award alone because it tests whether China can convert educational scale and patriotic recognition into a research environment that retains frontier talent.
The source is a commentary archive rather than a primary award release, so the social-media framing should be treated as selected evidence, not a representative poll.

Open at China Digital Times Ask Claude ChatGPT

A top Meta AI researcher leaves after one year to start a company

# Analysis QbitAI 21h

Chinese coverage focuses on Chinese-born researcher Jiahui Yu leaving Meta after roughly a year to found a stealth venture, highlighting the limits of even extraordinary AI compensation in retaining talent.

Yu joined Meta’s frontier-model effort after work at Google and OpenAI and helped lead multimodal projects including Muse Spark, voice, image and video systems.
His departure follows a recruitment campaign reportedly worth as much as $100 million annually, though the article acknowledges there is no authoritative confirmation of his personal compensation.
The move reflects a broader founder option for researchers who have reached the upper end of large-company pay, compute and organizational influence.
The article cites a database counting more than 200 notable researchers who left Meta and 929 former Meta researchers in total, but those figures are not independently verified here.
For Chinese technology observers, the story also represents a talent circuit in which researchers with Chinese educational backgrounds move through global labs before starting companies elsewhere.
The strategic implication for Meta is not simply one departure: expensive recruiting can build a team quickly, but it does not guarantee long-term ownership of the research agenda.

Open at QbitAI Ask Claude ChatGPT

Huawei says its atomic-service framework cuts cross-platform app work

# AI QbitAI 8h

Huawei is positioning HarmonyOS atomic services as reusable, context-aware service cards, with an enterprise framework that reportedly lets developers reuse existing assets instead of rebuilding apps.

The Atomic Service Cross Framework is designed to let developers carry existing mini-program and service assets into HarmonyOS with ecosystem-specific adaptation rather than a full rewrite.
The article says JD.com decomposed shopping functions into more than 200 contextual cards and that an AI coding plugin fixed 121 code anomalies during migration.
The platform model replaces app-first navigation with system-distributed cards, prompts and dialogue-triggered skills, giving Huawei more control over when a service appears.
A parking deployment across 11 Shenzhen malls illustrates the intended loop: contextual reminders, a persistent service card and small-value payment without repeatedly opening separate apps.
The strategic tradeoff is ecosystem dependence: lower migration cost can attract developers, while Huawei gains a system-level service graph and a larger role in transactions.

Open at QbitAI Ask Claude ChatGPT

DJI gets another chance to challenge US military blacklist

# Policy SCMP 16h

A US appeals court sent DJI’s challenge back for reconsideration without removing the Chinese drone maker from the Pentagon blacklist, keeping a major compliance fight alive.

The court did not clear DJI; it returned the case to a lower court to reconsider one part of the company’s challenge to its alleged military links.
For DJI, the designation can affect government procurement, investor perception and counterparties even if commercial drone sales continue outside the US defense market.
The case illustrates how Chinese technology firms face several US restriction mechanisms that are distinct from export controls and can be contested through administrative litigation.
A remand is procedural rather than a substantive victory, but it creates another opportunity to challenge the evidentiary basis for the listing.
The strategic question is whether the litigation produces a narrower designation standard or simply prolongs uncertainty for Chinese hardware companies operating globally.

Open at SCMP Ask Claude ChatGPT

SenseTime proposes Tokens Per Watt as an AI infrastructure metric

# Signal Leiphone 23h

SenseTime’s data-center unit argues that AI infrastructure should optimize effective Tokens Per Watt, not just PUE, and reports a roughly 60 percent increase in usable IT capacity through coordinated power and compute scheduling.

PUE measures facility overhead relative to IT power, but it does not say how much useful model output a watt produces; TPW links electricity, compute utilization, task value and time-of-use pricing.
SenseTime describes an eight-level observability layer mapping jobs and GPUs to racks, buildings and incoming power, so operators can see which workloads consume electricity and whether they can move.
Real-time inference stays fixed while training, evaluation and batch workloads can shift toward cheaper or less constrained periods, increasing output from existing capacity rather than simply adding megawatts.
The company says its Lingang AIDC improved effective IT carrying capacity by about 60 percent under existing power conditions, but the excerpt does not provide an independent audit or baseline.
The metric is strategically important if cloud customers begin paying for useful token throughput rather than accelerator-hours, forcing model efficiency, scheduling and energy procurement into one optimization problem.
TPW could become a useful commercial language, but it will need standardized definitions of effective tokens, quality, latency and workload comparability before it can support cross-provider benchmarking.

Open at Leiphone Ask Claude ChatGPT

DeepSeek Harness becomes a community-built agent platform

# AI QbitAI 7h

Chinese developers have created more than 700 public DeepSeek Harness plugins, extending the open-source agent shell with memory, vision, databases, games and multi-agent teams.

The harness uses an everything-is-a-plugin architecture, allowing community modules to add tools without changing the core product.
Examples include multi-agent team orchestration, an integrated file and terminal sidebar, file references, image understanding and persistent project memory.
The breadth of plugins shows that an agent platform’s value can shift from the base model to an ecosystem of composable tools and user-maintained workflows.
The article is enthusiastic and does not establish plugin security, compatibility or maintenance quality, which are the costs of an open extension surface.

Open at QbitAI Ask Claude ChatGPT

Why Chinese young people are turning to an elderly Cambridge professor

# Culture SCMP 2h

An 84-year-old British anthropologist has become an online confidant for Chinese youth, a small but revealing signal about loneliness, trust and the search for non-commercial advice.

Alan Macfarlane’s videos answer questions from Chinese viewers about friendship and other personal difficulties, using a calm, conversational style rather than influencer spectacle.
The appeal suggests that some young Chinese users are looking beyond domestic platforms and peers for emotionally credible guidance.
An elderly foreign academic can appear safer and less entangled in local status competition than a celebrity, career coach or algorithmically optimized influencer.
The phenomenon is not evidence of a national trend by itself, but it exposes a social demand for companionship and interpretation that mainstream productivity culture does not satisfy.
Its importance lies in how social platforms let a distant individual occupy an intimate advisory role in Chinese online life.

Open at SCMP Ask Claude ChatGPT

Chinese magnetic sensor could detect submarines from a smartwatch

# Analysis SCMP 9h

Chinese Academy of Sciences teams report a Hall-effect magnetic sensor sensitive enough to support future wearable detection of a submarine 500 meters below, though the consumer scenario remains speculative.

The research comes from institutes in Hefei and Ningbo and concerns magnetic-anomaly detection rather than acoustic sonar or radio transmission.
A sensor that could detect weak low-frequency magnetic fields at wearable size would have uses in navigation, industrial inspection and defense in addition to maritime monitoring.
The smartwatch scenario is an extrapolation from a laboratory materials result, not a demonstrated field product; the item provides no range test, false-alarm rate or wearable prototype data.
The story matters because it points to a sensor architecture where high sensitivity and low power, rather than compute scale, determine whether the technology can leave the lab.

Open at SCMP Ask Claude ChatGPT

WorkBuddy tries to make AI-generated HTML maintainable

# AI GeekPark 6h

WorkBuddy is adding collaborative editing, version history and structured data to AI-generated HTML, addressing the gap between producing a page and operating it as a durable work tool.

The article identifies local browser storage as the common failure mode: a generated HTML workspace can look complete but lose data across devices or after a cache reset.
WorkBuddy’s update lets users edit pages visually or in code, ask AI to modify a selected region, collaborate in a shared space and publish an updating link.
Its lightweight-app model separates HTML interaction from smart CSV data, an architectural move toward persistent dashboards rather than disposable generated pages.
Version history records human and AI changes, giving nontechnical users an approval and rollback mechanism that ordinary one-shot code generation lacks.
The product is testing whether AI-generated interfaces can become maintained organizational assets, although complex permissions, sensitive data and deep interactions still require conventional development controls.

Open at GeekPark Ask Claude ChatGPT

Baidu turns its GenFlow agent into a standalone office product

# AI Leiphone 1d

Baidu is separating its document-and-drive agent into a standalone office product after reporting more than 25 million AI-office monthly users, using private documents, memory and enterprise controls as its moat.

The product combines Baidu’s document corpus, academic content and user cloud files with Office agents that can generate and edit presentations, spreadsheets, documents, web pages and other media.
Baidu says GenFlow had more than 100 million monthly active users in April and that AI-office monthly active users exceeded 25 million; these are company-reported figures.
The enterprise version emphasizes shared context, task handoff, reusable skills, identity controls and full audit trails, addressing the organizational governance gap between chatbots and business systems.
A finance-focused mode packages company data, research reports and spreadsheet skills into workflows for valuation, research and meeting records, showing a verticalized route to monetization.
The separation from the storage product is strategically notable: the file repository becomes persistent memory and proprietary context for an agent, rather than merely a place to save generated files.
The claims of industry-leading evaluation scores and broad adoption need independent validation, but the architecture reflects where enterprise agents are heading: data access, tools, memory and governance in one surface.

Open at Leiphone Ask Claude ChatGPT

China condemns Japanese officials’ Yasukuni Shrine visit

# Policy SCMP 6h

Beijing lodged a formal protest after Japanese officials visited Yasukuni Shrine on the anniversary of Japan’s wartime surrender, keeping historical memory central to China-Japan diplomacy.

China described the visit and offering by Japanese defense chief Shinjiro Koizumi and others as unacceptable because many Asian countries associate the shrine with Japanese militarism.
Prime Minister Sanae Takaichi did not attend in person but reportedly sent a ritual offering, creating a distinction between direct and symbolic participation.
The response shows how wartime memory remains an active state-to-state signaling channel, especially when security tensions are already elevated.
The item does not report a new policy measure, so its immediate effect is diplomatic positioning rather than a material change in bilateral relations.

Open at SCMP Ask Claude ChatGPT

Viture bets lighter, cheaper XR glasses can create a real habit

# Analysis 36Kr 2h

Chinese XR maker Viture says its five-year strategy is turning display glasses into an everyday computer, using 63-gram weight, 1,999 yuan pricing (roughly $280) and built-in vision correction to widen adoption.

The company’s founder argues that earlier glasses were essentially external displays, while the product category must offer computing, interaction and content to justify another device on a user’s face.
The Pro 2 uses a Sony Micro-OLED display, claims 1,600 nits peak brightness and supports independent correction adjustments for near-sighted users.
Built-in correction is a significant China-market design choice in a country with roughly 600 million near-sighted people, according to the article’s framing.
The product’s early user feedback reportedly centers on lightness, clarity and comfort, with transit, rail travel and bed-side viewing emerging as practical use cases.
The deeper challenge is retention: low weight and lower price can improve trial, but a lasting category still needs frequent content or an agent-like function that makes wearing the device habitual.
Viture’s global sales claims and media reception come from the company-linked article, so they should be treated as early market evidence rather than independent market share data.

Open at 36Kr Ask Claude ChatGPT

US policy students are questioning the standard China-threat playbook

# Analysis SCMP 10h

Chinese analysis sees a generational shift among US foreign-policy students who are less willing to treat Beijing primarily as a strategic rival, a small but meaningful signal about future policy elites.

The article points to classrooms at Georgetown, Rice and other universities where students shaped by China’s rise are questioning an automatic move toward tougher policy.
This is an attitudinal trend, not evidence that US policy has changed; the current security and technology apparatus may remain more durable than classroom debate.
The signal matters because professional networks and assumptions formed during training influence the options future diplomats consider normal.
The Chinese framing emphasizes uncertainty inside Washington rather than presenting US-China competition as an irreversible consensus.

Open at SCMP Ask Claude ChatGPT

Chinese robot maker unveils a dual-arm machine with 50kg payload

# AI Leiphone 8h

Lumos Robotics has launched a wheeled dual-arm robot for industrial work, combining 50kg total arm payload, omnidirectional movement and force-aware sensing with a skill-training platform.

The MOS2 is about 1.65 meters tall, uses 22 degrees of freedom and combines a mobile base with two manipulators for tasks that fixed automation handles poorly.
Its sensing stack includes lidar, wrist cameras, force sensors and stereo vision, while hot-swappable batteries and a claimed five-million-cycle harness test target sustained industrial operation.
The company frames the product as an AI worker for flexible material handling, inspection and machine assistance rather than a humanoid demonstration platform.
Its NexCore software is intended to close a learning loop from real, simulated and video data through training, evaluation and deployment back into factory work.
The business test is not payload alone but task success, uptime, integration time and the ability to reuse skills across sites; the item provides no customer or production metrics yet.

Open at Leiphone Ask Claude ChatGPT

Sota Unbounded puts a physical world model into European supermarkets

# AI Leiphone 8h

A four-month-old Chinese startup says it will deploy more than 1,000 embodied robots with a major European supermarket group, betting that difficult retail environments can train general-purpose physical intelligence.

The company does not build robot bodies; it supplies an alleged native 4D world-action model intended to understand shelves, products, people and actions in real stores.
Its proposed path is domain convergence before broad generalization: solve a high-variation supermarket environment first, then transfer learned skills toward homes and industrial logistics.
The startup says it will collect up to 100,000 hours of initial physical-operation data and grow toward one million hours annually, with local storage, de-identification and privacy compliance required in Europe.
Supermarkets are a demanding intermediate test because inventory, layouts and human interactions change continuously, unlike tightly controlled factory cells.
The claims come from a company-led announcement and include no deployment metrics, task-success rate or customer name, so the commercial breakthrough remains prospective.
If the partnership is real and sustained, it would show a Chinese embodied-AI company using overseas enterprise access and compliance capability as a moat rather than relying only on domestic pilots.

Open at Leiphone Ask Claude ChatGPT

Chinese web novelist admits faking her death three years ago

# Culture IT Home 53m

A Chinese online novelist says a fake death announcement was used to keep her work circulating; readers accuse her of exploiting grief, and the platform is investigating.

In 2023, an alleged boyfriend announced the author’s death and said stored drafts and friends would continue the novel, after which the work reportedly gained traffic and tipping revenue.
The author later said she had survived a period of severe mental distress, but deleted the statement soon after posting it; the account and earlier chapter were also changed.
Readers now suspect the apparent return may relate to a newer facial-recognition requirement for withdrawals, which would make maintaining an absent author identity harder; that explanation remains online speculation.
The controversy shows how serialized fiction platforms turn personal authenticity, audience attachment and payment mechanics into one monetization system.
The platform said it would verify the author’s identity and act if rules were broken, illustrating how content moderation also becomes financial-account governance.

Open at IT Home Ask Claude ChatGPT

Nine convicted over 40,000 tonnes of adulterated diesel

# Policy Jiemian 5m

A Beijing case shows Chinese prosecutors pairing criminal convictions with environmental public-interest litigation after adulterated diesel caused documented air-pollution damage.

The defendants allegedly blended industrial white oil, crude white oil and fuel oil, then sold more than 40,000 tonnes as vehicle diesel for over 200 million yuan (roughly $28 million).
The court ordered more than 8.32 million yuan in compensation (roughly $1.2 million), while prosecutors treated sulfur and particulate emissions as a public environmental injury rather than only a consumer-fraud loss.
The case was discovered through criminal-prosecution work and transferred to a civil public-interest process, illustrating how Chinese procuratorates use environmental litigation to widen enforcement reach.
The same item says prosecutors handled 73,000 environmental public-interest cases nationwide in 2025, including 61,000 administrative and 12,000 civil cases.
Its broader signal is that product-quality enforcement can now bring state environmental institutions, financial evidence and pollution assessment into one case.

Open at Jiemian Ask Claude ChatGPT

Pikmin Bloom becomes a daily ritual in Taiwan and Hong Kong

# Culture Initium 13h

A location-based game has surged across Taiwan and Hong Kong, turning walking, flower planting and casual encounters into a shared routine and online social identity.

The game’s growth accelerated after late 2025, with social-media discussion reportedly rising sharply between February and April and monthly online mentions at one point growing 94.4 percent.
Players describe changing real-world behavior around walking goals, community days and digital traces left by people they pass in the street.
Its appeal differs from conventional competitive games: low-pressure routines, shared maps and collectible characters turn ordinary movement into social participation.
The article’s evidence concerns Taiwan and Hong Kong, so it should not be generalized to mainland Chinese gaming without additional data.
The story is useful as a culture signal about how location-aware games can shape public space and daily habits.

Open at Initium Ask Claude ChatGPT

Meituan says it will not open pharmacies, for now

# Analysis Jiemian 10m

Meituan again promises Chinese pharmacy operators it will stay out of physical retail, while industry coverage treats the pledge as a tactical move to secure supply and buy time.

Meituan executive Wang Puzhong said the company would not open physical pharmacies and would instead provide AI and digital services to the health-care sector.
The company already covers 250,000 pharmacies and 440,000 offline medical-consumption institutions, and has served 430 million users, giving it substantial leverage without owning stores.
The article argues that Meituan is expanding aggressively in other offline retail formats, so pharmacy operators view the promise as a phase-specific commercial choice rather than a permanent boundary.
In platform markets, controlling merchant supply can be more valuable than owning inventory; staying asset-light lets Meituan preserve that position while avoiding direct channel conflict.
The key test is whether regulation, margins or merchant dependence change enough to make vertical integration more attractive than continued intermediation.

Open at Jiemian Ask Claude ChatGPT

US may force partners to choose between rival AI frameworks

# Policy Techmeme 10h

US coverage says Washington plans to warn 35 partners that joining China’s competing framework could exclude them from Pax Silica, turning AI infrastructure alignment into an explicit bloc choice.

The reported policy would make participation in a US-led initiative conditional on avoiding a China-linked alternative, according to a Reuters source and letter.
This is an adjacent China story because its subject is US alliance policy, but China is substantive: standards, infrastructure and market access are being used as strategic instruments.
The approach could produce tighter interoperability within the US network while making neutral countries pay a diplomatic and technical switching cost.
It also risks accelerating the very bifurcation the policy seeks to manage, especially if Chinese frameworks offer cheaper deployment or access to markets that US partners value.
The item is an aggregator summary, so the proposed warning and the number of countries remain attributed claims rather than a confirmed public rule.

Open at Techmeme Ask Claude ChatGPT

China may lift travel restrictions on Manus founders as Meta deal unwinds

# Policy Techmeme 6h

US coverage says China may soon lift a travel ban on Manus founders as the Chinese AI company unwinds its $2 billion Meta acquisition; the story shows state leverage over cross-border AI transactions.

The report says Manus chief executive Xiao Hong plans to return to Singapore as the separation from Meta approaches resolution, citing unnamed sources.
The proposed lifting of restrictions would suggest that travel controls were part of the state’s leverage during a politically sensitive foreign acquisition, although the report does not explain the legal basis.
The deal’s unwinding matters beyond Manus: Chinese AI firms can face constraints not visible in ordinary merger announcements, especially when ownership, talent and strategic technology cross borders.
This is an aggregator summary of Financial Times reporting, so the travel-ban and timetable claims remain source-attributed rather than independently established.

Open at Techmeme Ask Claude ChatGPT

Shanghai’s tech index rises 29% as Beijing’s AI push fuels a frenzy

# Signal Techmeme 24m

US tech coverage highlights a 29% 2026 gain and a 150-plus P/E for Shanghai’s STAR 50, showing how Beijing’s technology push is feeding a high-risk domestic equity trade.

The reported valuation is far above the Nasdaq 100’s cited P/E of 35, making the index a measure of policy-backed expectations as much as current earnings.
The rally connects state industrial priorities with retail and institutional investor enthusiasm, a recurring feature of Chinese capital markets.
A high multiple can fund ambitious technology companies, but it also makes policy disappointment, weak monetization or tighter market controls more consequential.
The aggregator item is evidence of what US financial media is watching; it does not independently verify the underlying index data beyond the cited Financial Times report.

Open at Techmeme Ask Claude ChatGPT

China’s beauty industry is searching for a cultural premium

# Culture Jiemian 15m

Chinese beauty brands are moving beyond low prices and ingredient claims toward cultural identity, as a 1.1 trillion yuan market (roughly $155 billion) matures and growth slows.

The article says China has become the world’s largest cosmetics-consumption market, but brands face exhausted traffic, price competition and defensive discounting by international players.
Its proposed C-Beauty thesis treats cultural identity as the next growth engine after e-commerce distribution and efficacy-led products.
That positioning reflects a broader shift in Chinese consumer branding: founders are discussing aesthetics, heritage and a sense of place instead of only gross merchandise volume and advertising return.
The article links this supply-side change to state investment in museums, galleries and cultural heritage, which has made cultural references more available as everyday commercial material.
The strategic question is whether cultural storytelling can create durable global brands or becomes another short-lived marketing layer over commoditized products.

Open at Jiemian Ask Claude ChatGPT

Chinese tech roundup mixes AI releases with labor and consumer-policy signals

# Culture 36Kr 11h

A Chinese roundup combines routine product news with a notable retailer plan to hire former prisoners and a state campaign voiding 1,782 deceptive trademarks, offering a snapshot of changing commercial norms.

The retailer Pang Dong Lai plans to hire 20 people who served sentences of at least five years, alongside 100 retired border-defense personnel, framing employment as dignity and social reintegration.
China’s trademark authority has voided 1,782 registrations judged deceptive, including names implying rustic production, zero sugar, zero additives or misleading origin and efficacy.
Together, the two items show Chinese businesses and regulators contesting how trust is built: through inclusion and rehabilitation on one side, and tighter control of marketing claims on the other.
The roundup also repeats AI product and financial news, including claims about an open model release, but those snippets are not a substitute for the dedicated technical reports.
Because this is a multi-story roundup, its value lies in the social and regulatory signals rather than in any single product announcement.

Open at 36Kr Ask Claude ChatGPT

ChinAI	403	0/0		HTTPError: 403 Client Error: Forbidden for url: https://chinai.substack.com/feed
ChinaTalk	200	0/20
Pekingnology	200	0/20
Ginger River Review	200	0/20
Sinocism	200	0/20
QbitAI	200	9/10
36Kr	200	4/30
Solidot	200	2/20
Leiphone	200	8/20
GeekPark	200	2/30
IT Home	200	12/60	+47 over cap
SCMP	200	10/50	+3 over cap
Initium	200	3/15
Jiemian	200	8/30	+22 over cap
China Digital Times	200	2/15
The Wire China	200	1/50
Techmeme	200	4/15
Hacker News	200	1/20
V2EX	200	8/50	+11 over cap

The US answered China's memory boom by telling Apple no and handing 35 capitals a choice

Lutnick tells Apple to skip Chinese DRAM

Thirty-five capitals told to pick a side

Beijing's own exit control, quietly lifted

China's second legal code took effect today

A 27-billion model that fits a 3090

Wang Hong said "discouraged" and Weibo blinked

Qwen3.8-27B reads differently on Hacker News

Threads we are pulling