China Desk — 23 Aug 2026

Why “flexibility” sounds like an insult to Chinese workers

# Pekingnology

A Chinese economist’s claim that flexibility is a form of welfare triggered online backlash because gig workers heard it as praise for insecurity and weak social protection.

The controversy turns on a mismatch between labor-economics language and lived experience: autonomy can have value, but it is not equivalent to pensions, employer contributions or predictable income.
The backlash shows how precarious employment has become a political and emotional fault line in China, especially when elite speakers describe it in abstract welfare terms.
The professor’s fuller argument distinguishes time control from benefits, yet the viral sentence traveled without those qualifications and was compared online to elite indifference toward poverty.
For employers, the episode is a warning that flexible-work rhetoric will be judged against income volatility, platform power and the actual distribution of risk.
The broader policy question is whether China can extend social insurance to its enormous flexible-worker population without treating flexibility as a substitute for security.

Open at Pekingnology Ask Claude ChatGPT

Why banquet assault cases expose China’s power problem

# China Digital Times

Chinese coverage of alleged assaults by officials stresses censorship and power imbalance: reports can disappear or wait for official notices, while commentators argue case-by-case punishment is not enough.

The translated post says outlets risk being pressured or removed when local authorities have not issued an official statement, making official timing a gate on investigative reporting.
One official’s biography reportedly vanished from a government website while a newspaper article about the case disappeared, showing how administrative opacity can replace transparent status updates.
The author rejects an individual-morality explanation and frames banquet assault as a power-accountability problem involving officials and state-enterprise executives.
That framing matters because China often responds to scandals through disciplinary announcements that identify punishable individuals without exposing the institutional conditions that enabled them.
The strongest implication is not that every allegation is proven, but that censorship and selective disclosure make it difficult for the public to distinguish investigation, punishment and information control.

Open at China Digital Times Ask Claude ChatGPT

China's next five-year health plan calls for nationwide long-term care insurance

# Jiemian

China's 2026-2030 health plan commits to a nationwide long-term care insurance system, moving eldercare from local pilots toward a national framework.

The plan targets a multi-tier health-insurance system by 2030 and explicitly includes long-term care insurance as a major reform task.
The proposed system generally reimburses approved care services rather than paying cash directly, linking insurance expansion to the supply and quality of caregivers.
A national disability-assessment standard and mutual recognition across regions would address one of the main problems of fragmented local pilots.
The policy is demographic infrastructure: it shifts eldercare risk from families alone toward pooled financing and formal service providers.
The supplied analysis says the current service catalog contains 36 care and medical-care items, indicating that benefit design will be as important as coverage.
The central risk is implementation: national eligibility rules without enough trained caregivers, facilities or sustainable contributions could create nominal coverage rather than usable care.

Open at Jiemian Ask Claude ChatGPT

China's Zhuque-3 completes its first controlled land booster recovery

# IT Home

Chinese launch company LandSpace says Zhuque-3 successfully reached orbit and landed its first stage on land, turning recovery from a demonstration into a reuse-engineering program.

The August 19 mission is described as China's first successful controlled, landing-leg recovery of an orbital rocket stage on land.
The booster was transported back to the assembly facility for inspection, teardown of selected systems, condition analysis and maintenance before a planned reuse assessment.
The important milestone is not touchdown alone but the post-flight workflow needed to establish inspection criteria, refurbishment procedures and repeatable turnaround.
China has also demonstrated a separate sea recovery with a Long March booster, suggesting domestic launch providers are now testing multiple recovery geometries rather than one national prototype.
LandSpace says it plans to attempt a reflight within six months, but the supplied report does not establish whether the recovered hardware will meet that schedule.
If reuse becomes reliable, lower marginal launch cost and higher launch cadence would change the economics of China's commercial space sector; until reflight, the key uncertainty remains refurbishment depth and cycle life.

Open at IT Home Ask Claude ChatGPT

China recalls millions of EVs over hard-to-find emergency handles

# IT Home

Chinese coverage stresses a regulatory correction to minimalist EV design: nine automakers are recalling millions of cars because emergency mechanical releases can blend into the interior and delay escape.

The defect matters only when electronic door systems fail after a severe crash or immersion, but that is precisely when a purely mechanical release becomes the last independent escape path.
The reported fixes combine clearer physical warnings with software that can automatically lower windows after an accident, showing regulators accepting software remedies only alongside a physical fallback.
The source links the common defect to hidden handles, touch-heavy cabins and styling decisions that made emergency controls visually disappear.
A new mandatory national standard is scheduled for January 2027 and will require obvious mechanical opening structures and emergency markings on new cars.
The recall demonstrates a broader regulatory pattern in China: permit aggressive product experimentation, then use coordinated recalls and prospective standards to impose a common safety baseline.
The source does not establish that every recalled model has caused an accident; the finding is a design and discoverability hazard, not a reported failure count.

Open at IT Home Ask Claude ChatGPT

Qin Weiguo named Communist Party chief of Hefei

# Jiemian

Anhui has appointed Qin Weiguo, a provincial standing committee member, as Communist Party secretary of Hefei, a key change in the leadership of a major technology and manufacturing hub.

The appointment was announced by the Anhui Party organization and reported through the provincial official account.
Hefei is an important center for Chinese advanced manufacturing, electric vehicles, displays, and research, so its Party secretary has influence over industrial policy and local capital allocation.
The brief report gives no information about Qin's priorities, predecessor, or whether the move signals a change in Hefei's technology strategy.
The appointment is nevertheless worth tracking because local Party leadership often coordinates the state support, land, financing, and talent policies behind industrial clusters.

Open at Jiemian Ask Claude ChatGPT

Chinese coverage casts Nvidia's Poolside deal as a China-model challenge

# IT Home

Chinese coverage presents Nvidia's $6 billion Poolside deal as a push against DeepSeek and Kimi, while US coverage emphasizes building an open-weight American AI ecosystem.

The reported structure is a $6 billion model-software authorization plus a possible $1 billion investment at a reported $12 billion valuation; Poolside remains independent and the license is nonexclusive.
Nvidia would also offer roles to Poolside's 109 employees, indicating a hybrid strategy that combines licensing, talent access and capital rather than a conventional acquisition.
The Chinese framing makes the strategic target explicit: Chinese open models are treated as competitive benchmarks, not merely domestic curiosities.
The US framing in the paired coverage emphasizes ecosystem construction and Nvidia's incentive to expand inference demand for its chips.
The important architectural question is whether Nvidia can turn model access and hardware distribution into a durable developer platform, rather than merely sponsor another high-cost frontier effort.
The source attributes the deal terms to anonymous sources and a US newspaper, so the authorization, investment and valuation remain reported rather than company-confirmed facts.

Open at IT Home Ask Claude ChatGPT

A humanoid robot rallies with a tennis champion in a live showcase

# AI QbitAI

Galaxy General used a globally broadcast match to show a humanoid robot sensing, moving, striking, and coordinating with a human in an open-ended sport rather than following a fixed trajectory.

The demonstration covered multiple strokes, movement patterns, singles and doubles situations, and real-time choices about shot placement and positioning.
The company says its AstraBrain model unifies task reasoning, whole-body control, and low-level neural control instead of passing commands through several separately trained modules.
Tennis is a demanding systems test because the robot must estimate a fast ball, move its center of mass, coordinate its legs and torso, and control the racket before the next return arrives.
The reported training approach combines ordinary human motion fragments with a large virtual tennis environment, reducing dependence on expensive full-match motion capture.
The result is best understood as evidence that the company can close a perception-to-action loop in one narrow but difficult domain, not as proof of general household or industrial autonomy.
Chinese coverage casts the match as an analogue of a landmark game-playing breakthrough, a framing that is useful for understanding the sector’s ambition but not a substitute for independent benchmarks.

Open at QbitAI Ask Claude ChatGPT

A young Chinese startup bets on an end-to-end robot brain

# AI QbitAI

A team of doctoral researchers is building a full-body control model for bipedal robots, arguing that modular brain-and-motor stacks create information bottlenecks that limit scaling.

The company’s public demonstration used a robot to drive a go-kart through turns, requiring coordinated vision, hands, feet, balance, and force control at speed.
Its founder argues that separating high-level reasoning from motor control introduces translation losses and cascading errors, even when each module performs well in isolation.
The team is pursuing an end-to-end model that learns perception, task understanding, and action together, a more aggressive path than the modular architecture still favored for many production systems.
The technical difficulty is higher for a biped because reaching, bending, and contact change the robot’s center of mass and require whole-body dynamics, not just arm trajectories.
The company says its work spans perception, decision-making, control, data, and systems, with more than 40 conference papers and several claimed field firsts.
The central uncertainty is whether a single model can achieve reliable safety, interpretability, and transfer across tasks better than a carefully engineered hierarchy.

Open at QbitAI Ask Claude ChatGPT

A Beijing robot cafe tests whether embodied AI can actually work

# AI QbitAI

A robot cafe at Beijing’s World Robot Conference lets machines serve, clear tables, avoid people and belongings, and coordinate long task sequences in an open public space.

The service robot reportedly navigates around moving visitors, places drinks beside private objects, collects cups and tissues, disposes of waste, and sanitizes its gripper.
Unlike a staged booth, the cafe has customers ordering at will and changing the environment, making perception, replanning, semantic object handling, and recovery part of the demonstration.
The task chain combines navigation, manipulation of soft objects, human interaction, and long-horizon planning, which are usually shown as separate benchmarks.
The deployment was reportedly tested in a commercial cafe before appearing at the exhibition, a stronger signal than a fixed route but still not evidence of unattended operation at scale.
The Chinese company behind the system positions a shared general-purpose brain and body platform against robots optimized for one isolated action.
The report also includes company claims about European certification, financing, and orders; those claims are not independently substantiated in the supplied text.

Open at QbitAI Ask Claude ChatGPT

Developers hunt for the company behind anonymous Ox Alpha

# AI QbitAI

Community testing points toward Z.ai as a possible source of the anonymous Ox Alpha model, while Google remains an alternative; tokenizer, video, and API clues are suggestive but not proof.

Also: V2EX

Ox Alpha appeared on OpenRouter with a roughly one-million-token context window, text, image, and video input, and a large maximum output, making it attractive for long coding agents.
Testers reported that its token counts showed a fixed relationship with GLM-family models, possibly explained by a hidden system prompt.
Video-token behavior reportedly resembled a GLM multimodal service more closely than several competing models, a stronger clue because it reflects input-pipeline engineering rather than writing style alone.
Community members also found API error patterns and alleged a model-output trace pointing to Z.ai, but the report explicitly treats that evidence as weaker and secondhand.
Google remains plausible because the product profile resembles Gemini’s long-context, multimodal, agent-oriented positioning and because Google researchers made ambiguous social-media comments.
The identity has not been confirmed by OpenRouter or either candidate, so the episode is best read as an informal model-forensics exercise, not an attribution.

Chinese developers speculate about the identity of Ox Alpha V2EX

Open at QbitAI Ask Claude ChatGPT

Qiyuan opens consumer preorders for two personal robots

# AI QbitAI

Chinese materials company Shangwei New Materials has opened preorders for the Qiyuan Q1 and T1, moving its personal robots from demonstrations into an early test of household demand.

The Q1 is described as a small full-body force-controlled humanoid, while the T1 can switch between wheeled humanoid and quadruped forms.
The company says its robot business recorded 210 million yuan in advance payments, roughly $30 million, in the first half of 2026 and has experience stores in seven major cities.
It is also building a production base aimed at a first batch of about 10,000 units and has signed component and manufacturing partnerships.
The commercial question is not whether consumers will preorder a novel robot but whether the machines deliver repeated value in unstructured homes after the novelty fades.
Household robots face harder requirements than industrial systems because every home has different layouts, objects, people, and safety expectations.
The preorder is therefore a market-validation event, not evidence that general-purpose home robotics has reached mass adoption.

Open at QbitAI Ask Claude ChatGPT

A physical-AI model learns new robot tasks from one demonstration

# AI GeekPark

Generalist’s GEN-1.5 reports 59 percent average success on ten new tasks after a single physical demonstration, pointing toward reusable robot skills without parameter updates.

The model can infer a new way to use an object, such as sweeping with a banana or adapting when an obstacle appears, rather than replaying a fixed trajectory.
Its physical-prompting interface uses a three-to-12-second demonstration to specify the task while relying on an existing base model for grasping, contact and control.
Generalist says GEN-1.5 learned from more than 500,000 hours of physical data and pushed some earlier task averages from 64 percent to 99 percent across model generations.
The one-shot result is still weak for unattended work: five minutes of data and ten parameter updates raise the reported average to 83 percent, indicating that adaptation has not disappeared entirely.
The architectural bet is that physical intelligence can scale like language models, with broad data and compute producing transfer across tools and situations instead of a new training run for every task.

Open at GeekPark Ask Claude ChatGPT

The investor who backed Unitree now chairs another robot startup

# Analysis QbitAI

Investor Yin Fangming supplied Unitree’s first reported angel funding and now chairs Galaxy General, illustrating how China’s robotics ecosystem is shaped by repeat founders, investors, and hardware specialists.

The report says Yin invested 2 million yuan, roughly $280,000, in Unitree in 2016 for 15%, later receiving about 280 million yuan, roughly $39 million, in estimated returns after transfers.
He previously helped found ROOBO, a robotics and hardware incubator that invested in more than ten companies and also experienced the cost and failure modes of consumer hardware.
The article credits that background with helping him recognize early hardware potential in Unitree and later move into embodied-intelligence leadership.
Unitree’s reported strength in proprietary motors, reducers, controllers, and cost control contrasts with ROOBO’s earlier struggle to turn a broad hardware portfolio into durable demand.
This is a useful map of China’s industrial network: people move between startup building, angel investing, funds, and board roles as the robotics sector matures.
The return figures are the article’s calculations, not a current audited valuation or a complete accounting of Yin’s economic interest.

Open at QbitAI Ask Claude ChatGPT

A Chinese humanoid robot plays fully autonomous tennis

# AI Leiphone

Galaxy General’s robot completed a reported first fully autonomous humanoid tennis match, combining millisecond-level ball tracking, whole-body control and tactical adaptation.

The system must solve perception, footwork, torso rotation, racket timing and balance in one closed loop, rather than executing a scripted strike from a fixed position.
Galaxy General describes AstraBrain as a unified stack in which a higher-level model chooses tactics, a lower-level model controls the body and a bridge translates decisions into motor commands.
The team trained from fragmented ordinary-player motion data and used latent-space action constraints to keep generated movements within human-like ranges while adapting to the incoming ball.
A virtual tennis environment supplied large-scale practice against simulated opponents before transfer to the physical robot, reducing the cost of collecting complete professional matches.
The demonstration is a meaningful integration test, but a match against a human is not evidence of reliable industrial manipulation; repeatability, maintenance, safety and performance outside tennis remain the harder milestones.

Open at Leiphone Ask Claude ChatGPT

Wujie Power pitches hidden-space world models for real robots

# Analysis QbitAI

At Beijing’s World Robot Conference, Wujie Power argued that robots should learn physical causality in latent space and use industrial and commercial deployments as a continuous training loop.

The company contrasts latent-space prediction with pixel-level video generation, saying a useful world model should predict how actions change the environment rather than merely render the next frame.
Its proposed stack combines high-dimensional representations with reinforcement learning so robots can learn both physical regularities and action values.
The company’s deployment thesis is to use industrial work for reliable skills and commercial environments for generalization across lighting, layouts, objects, and crowds.
That makes deployment part of model development rather than a final integration phase: real operations generate data that feeds later policy updates.
The approach is a company position inside an unsettled field, not an established consensus; the report gives no independent benchmark against competing world-model architectures.
Wujie’s participation in standards and industry forums shows how Chinese robotics firms are trying to shape the vocabulary and norms of an emerging category.

Open at QbitAI Ask Claude ChatGPT

China delays Chang’e-7 lunar ice mission

# Analysis SCMP

China postponed Chang’e-7, a multi-spacecraft mission intended to study and drill for ice near the lunar south pole, after a comprehensive assessment by the human-spaceflight authority.

Also: IT Home, Jiemian

The mission combines an orbiter, lander, rover and hopping vehicle, making it a systems test for operating across permanently shadowed and difficult south-polar terrain.
The delay is strategically important because lunar polar volatiles are central to future in-space resource and sustained-surface-operations plans, not merely a science target.
The source gives no new launch date or technical cause, so it would be wrong to infer whether the delay is a minor schedule adjustment or a deeper vehicle problem.
China’s public disclosure is terse but consistent with a program culture that prioritizes mission assurance over maintaining a previously expected launch window.

China says Chang'e-7 misses this year's launch window IT Home
Chang'e-7 will not launch in its planned window Jiemian

Open at SCMP Ask Claude ChatGPT

Galaxy General wants one shared brain across robot bodies

# AI GeekPark

At Beijing’s robot conference, Galaxy General showed a biped, wheeled dual-arm and heavy-load robots sharing an upper-layer perception, planning and world-model stack.

The architecture does not make bodies interchangeable: each platform retains its own dynamics and low-level control, while perception, task understanding and planning are intended to transfer across forms.
The new ET1 biped uses the AstraBrain-WBC model to map observed human motion into whole-body actions, drawing on about 100,000 hours of motion data.
This is a platform strategy rather than a single humanoid launch: the company wants experience gathered by one body to reduce the data and engineering burden for the next body.
The hard question is whether shared representations actually improve deployment economics without forcing every customer scenario through a lowest-common-denominator control layer.

Open at GeekPark Ask Claude ChatGPT

Humanoid robots outrun people but struggle to stop

# AI Solidot

At Beijing’s humanoid robot games, two robots reportedly beat human sprint records, while several crashed into barriers because braking and recovery remain weak.

The event includes 51 contests, with more than 40 percent requiring fully autonomous operation and additional tests in factories, restaurants, offices and emergency scenarios.
The performance gap is instructive: fast locomotion can be demonstrated in a controlled straight line, while braking, balance recovery and mechanical durability expose the harder systems problem.
The organizers are testing mundane tasks such as cable connection and displaced packages, which better approximate deployment than headline athletic demonstrations.

Open at Solidot Ask Claude ChatGPT

China builds programmable plant immunity for outbreak response

# AI SCMP

Chinese scientists report an AI-guided platform that designs synthetic plant immune receptors to detect pathogen proteins and trigger targeted defenses against crop diseases.

The receptors are programmable biological sensors: researchers can design them to recognize proteins from particular bacteria, viruses or fungi and activate a plant immune response.
The approach could shorten the response cycle for emerging crop diseases compared with breeding or deploying broad-spectrum treatments, but it still faces delivery, field-environment and ecological validation problems.
Its strategic value is agricultural resilience: a reusable design platform could protect commercial crops without waiting for a conventional resistance-breeding pipeline.
The source calls it the world’s first custom plant-immunity system, but provides no field-trial results, crop coverage or quantified disease-control improvement.

Open at SCMP Ask Claude ChatGPT

A Chinese roundup tracks humanoid races, DeepSeek pricing and a mystery model

# GeekPark

A Chinese roundup leads with Beijing’s humanoid robot games, DeepSeek’s weekend low-price API tier and an anonymous model claimed to beat named rivals on code repair.

The robot competition drew 666 teams and 2,056 machines across 51 events, while the roundup presents the event as a test of autonomy in industrial, logistics and emergency scenarios rather than only athletic spectacle.
DeepSeek’s new weekend pricing makes its low-traffic period uniformly cheaper; the cited flagship output price falls from 27 yuan to 13.5 yuan per million tokens, roughly $3.80 to $1.90.
The anonymous Ox Alpha model reportedly scored about 80 percent on ten DeepSWE tasks, but the source itself notes that its provenance is unclear and the comparison is not an independent benchmark.
The item also reports Anthropic hiring a former Google TPU leader and Tesla preparing a Cybercab event, making it a domestic attention map rather than a single verified event.

Open at GeekPark Ask Claude ChatGPT

Why Chinese media says US historians arm separatist narratives

# Policy SCMP

Chinese state-linked coverage attacks a US-born interpretation of Qing history as theoretical ammunition for separatism, folding academic debate into ethnic policy and national-unity politics.

The dispute shows how China’s official historical narrative extends beyond textbook content into debates over whether the Qing should be understood through imperial, ethnic or modern national frameworks.
Calling a scholarly approach separatist ammunition shifts the issue from historiography to political security, narrowing the space for interpretations that challenge the state’s preferred national story.
The timing also reflects growing official concern over historical nihilism and online pseudo-history, where competing accounts can become proxies for present-day identity politics.

Open at SCMP Ask Claude ChatGPT

Beijing’s willow and poplar trees may worsen ozone pollution

# Signal Solidot

A study reports that plant emissions supplied about 10 percent of Beijing’s volatile-organic-compound emissions but 52 percent of the chemicals driving ozone formation during the study period.

About 35 percent of Beijing’s trees reportedly emit isoprene, which reacts with hydroxyl radicals and nitrogen oxides under sunlight to generate ozone-forming compounds.
The result complicates the simple idea that more urban greenery always improves air quality: species selection and atmospheric chemistry can matter as much as tree count.
The study used Beijing air samples and ozone observations from May through July 2021, so its seasonal and city-specific limits should not be ignored.

Open at Solidot Ask Claude ChatGPT

African companies adopt Chinese AI while keeping US options open

# Analysis SCMP

Businesses in Africa are using Chinese models where cost or flexibility is better, suggesting AI markets may remain pragmatic and multi-polar despite US pressure to choose sides.

The account describes Nigerian insurtech Curacel expanding its model stack beyond Silicon Valley, treating Chinese systems as a commercial option rather than a geopolitical declaration.
The competitive advantage for Chinese providers is presented as price and adaptability, not necessarily superior frontier capability; procurement decisions will vary by data policy, support and performance.
For China, overseas developer adoption offers a route to influence through software deployment even when chip and cloud access remain contested.

Open at SCMP Ask Claude ChatGPT

Researchers find Android malware targeting a Chinese car head unit

# Signal Solidot

Kaspersky says malware spread through the built-in updater of an Android-based Chinese vehicle head unit, apparently to recruit cars into a residential proxy botnet.

The attack resembles cheap-TV-box infections: compromised devices become network infrastructure for attackers rather than merely displaying malware to the vehicle owner.
Kaspersky says it notified the automaker and that the security issue was fixed, but the report does not identify the vulnerable software component or patch verification method.
The incident shows why vehicle infotainment systems need a software-supply-chain model closer to phones and smart TVs, including signed updates, component inventory and post-sale patch governance.

Open at Solidot Ask Claude ChatGPT

Chinese scholar says US AI supply-chain strategy will backfire

# Policy SCMP

A Chinese political scientist argues that US chip, rare-earth and AI alliances will fail and that restricting access to low-cost Chinese models could accelerate the very rivalry they seek to contain.

The argument reflects a Chinese policy frame in which supply-chain alliances are portrayed as exclusionary and economically self-defeating rather than as durable security coalitions.
The scholar’s proposed alternative is US-China cooperation to expand global demand, revealing how some Chinese economic commentary treats external growth as a prerequisite for long-term stability.
This is an attributed opinion, not evidence that the initiatives will fail; its value is showing how Beijing-adjacent debate interprets containment and model competition.

Open at SCMP Ask Claude ChatGPT

Chinese officials moonlight as delivery riders in a test of regulation

# Culture Jiemian

Binzhou sent 10 young officials to work as delivery riders for two weeks, collecting 145 problems and promising a closed loop from field observation to rectification.

The program required officials to operate like regular riders rather than arrive as identifiable regulators, because workers often withhold complaints when officials are present.
Assignments covered business licenses, pickup locations, kitchen hygiene, waste and package seals, turning immersion into sector-specific regulatory data collection.
The experiment is unusually explicit about the risk of performative fieldwork: its credibility depends on whether the reported findings produce measurable changes after the publicity fades.
It also reveals a governance pattern in which cadres are asked to experience platform labor directly to compensate for the distance between administrative categories and workers' daily reality.

Open at Jiemian Ask Claude ChatGPT

A Chinese roundup mixes rumor arrests, graduate wages and AI pricing

# Culture 36Kr

Chinese coverage highlights an arrest over false IPO rumors, vocational graduates out-earning some degree holders, DeepSeek’s weekend discount and Alibaba’s AI financing.

Police detained the author of a false claim that Xiaohongshu’s IPO had failed, illustrating how online financial rumor-making can become an administrative-order case rather than merely a platform moderation issue.
The roundup cites research showing some high-paying vocational fields, including rail transport and power engineering, exceeding the earnings of lower-paid university majors six months after graduation.
DeepSeek’s weekend API pricing change and Alibaba’s planned 80 billion Hong Kong dollar placement show Chinese AI competition operating through both cheaper usage and very large infrastructure commitments.
As a digest, the item is more useful for mapping domestic attention than for independently establishing each underlying claim.

Open at 36Kr Ask Claude ChatGPT

Washington takes an America First line in APEC talks in China

# Policy SCMP

US officials are promoting an America First agenda at APEC meetings hosted in China weeks before Xi Jinping’s planned US visit, making trade language part of the diplomatic runway.

The US framing emphasizes reciprocal trade, removing barriers for American business and economic growth, while the setting places those demands inside a China-hosted multilateral forum.
The timing matters because APEC can provide procedural contact even when bilateral technology and trade relations are being securitized.
The report gives no Chinese response, so it shows Washington’s public position more clearly than the likely negotiating balance.

Open at SCMP Ask Claude ChatGPT

History exam changes may leave Chinese students open to pseudo-history

# Culture SCMP

A Chinese academic warns that students dropping history as an exam subject could weaken public resistance to distorted narratives as Beijing intensifies its campaign against historical nihilism.

The concern is not simply lower historical knowledge: it is that reduced formal study may make online audiences more dependent on simplified nationalist stories, viral myths or state-approved explanations.
The debate shows education policy, historical scholarship and political legitimacy intersecting in China, where the acceptable national narrative is treated as a governance issue.
The report does not establish how widespread the subject-choice shift is or measure its effect on public beliefs, so the warning remains an informed forecast.

Open at SCMP Ask Claude ChatGPT

Pony AI doubles revenue but remains far from profitability

# Signal Jiemian

Pony AI reported first-half revenue of $70.47 million and a 534% jump in Robotaxi revenue, yet its net loss widened to $98.86 million, leaving scale economics unresolved.

The report says total revenue rose 98.9% year on year and Robotaxi revenue reached $20.6 million, but first-half net loss still expanded 9.1%.
Its net-loss ratio fell from 255.8% to 140.3%, a useful growth-stage metric but not evidence of breakeven; gross margin remained below 17%.
The article notes that Robotaxi vehicles in Guangzhou and Shenzhen have reportedly reached positive unit economics, while the consolidated accounts still reflect heavy expansion costs.
The strategic question is whether fleet density, utilization and lower operating cost can convert local unit economics into company-wide profit before investors lose patience.

Open at Jiemian Ask Claude ChatGPT

ByteDance's free fiction platform takes aim at China Literature's moat

# Culture Jiemian

A blockbuster author's move to ByteDance's free-reading platform exposes a structural fight with China Literature: algorithmic distribution and short-video IP are challenging the paid web-fiction model.

The report says the author behind Battle Through the Heavens launched exclusively on Tomato Fiction, drawing more than 10 million readers within weeks; these figures come from the outlet and company-linked reporting.
QuestMobile data cited by the article puts Tomato Fiction at 260 million monthly active users in June, versus 134.1 million for China Literature's platforms, whose paying users fell 10.9% year on year.
The article estimates the wider Tomato ecosystem generated more than 600 billion yuan in 2025, roughly $84.5 billion, while China Literature reported 35.3 billion yuan in first-half revenue, roughly $5.0 billion.
The strategic divide is free, algorithmic distribution backed by ByteDance against paid reading, rights libraries and long-cycle intellectual-property development.
The deeper threat is production economics: the article argues that generative AI could let the platform with the larger behavioral data loop iterate stories and adaptations faster, although it provides no independent measurement of that advantage.

Open at Jiemian Ask Claude ChatGPT

Chinese archaeologists begin a joint project in Peru

# Culture SCMP

China’s archaeology institute will work with Peruvian partners at ancient sites, marking its first joint field project in South America and extending Beijing’s cultural ties.

The partners include Peru’s Caral archaeological authority and the National University of San Marcos, linking Chinese research institutions with a major Peruvian heritage site.
The project is cultural diplomacy in a slower and less transactional form: shared fieldwork creates institutional relationships and a narrative of civilizational exchange.
The supplied report does not yet provide excavation findings, funding figures or a timetable for discoveries.

Open at SCMP Ask Claude ChatGPT

China orders rapid investigation into formaldehyde-treated cabbage

# Policy Jiemian

Three national agencies ordered an urgent investigation and nationwide inspections after reports that cabbage was dipped in formaldehyde during purchase, showing how food scares trigger centralized enforcement.

The State Council food-safety office, agriculture ministry, and market regulator directed local authorities to trace the affected cabbage and punish illegal conduct.
The agencies also ordered broader inspections of perishable vegetables, turning one local allegation into a national compliance exercise.
The episode shows China's food-safety system combining rapid media-driven response with administrative campaigns that seek to demonstrate control before the facts are fully settled.
The report does not establish how widespread the practice was or identify final penalties, so the immediate significance is the state response rather than the scale of contamination.

Open at Jiemian Ask Claude ChatGPT

Chang'e-7 slips a year, and only Thailand said why

The moon window closes for the year

A labour economist becomes Marie Antoinette

Nobody will admit to Ox Alpha

Robot records fall, buyers ask about payback

History gets treated as a security problem

Threads we are pulling