China Desk

China News, Summarized 24 Aug 2026, 04:56 UTC 34 stories 18/19 sources

Today's brief

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

The moon window closes for the year

The mission is off. Sunday evening in Beijing, the crewed space programme office told CCTV that Chang'e-7 does not meet launch conditions and cannot fly in this year's window, invoking the principle of steadiness and leaving nothing to chance. The probe reached Wenchang in April, its Long March 5 in July; both had finished assembly and testing.

Nobody in Beijing mentioned weather. Thailand's National Astronomical Research Institute, which built one of the mission's instruments, said the postponement followed Tropical Storm Narra forming in the South China Sea, and noted that the south-pole target makes for a very particular window in which the sun lines up on the crater.

Chinese wires ran that storm as a wave bulletin. The national marine forecast office warned of 11-to-19-foot seas in the Beibu Gulf from the season's 19th typhoon. Wenchang is on Hainan, under the same alert. Same storm, two stories, and the foreign partner is the one who joined them up.

That is how a week costs a year. Landing was set for November after two months in lunar orbit; observers had the window closing 31 August. Chang'e-7 was to be the first mission to land directly at the pole and hunt ice above 85 degrees south. With Chang'e-8 around 2028 and a crewed landing pencilled for 2030, there is less slack here than the decade suggests.

Go deeper on this section: ClaudeChatGPT

A labour economist becomes Marie Antoinette

Zhang Dandan said flexibility is welfare. The Peking University economics professor and deputy dean of its National School of Development, asked on a talk show about social protection for gig workers, replied that in labour economics flexibility is itself a form of welfare — salaried staff trade freedom for pensions; flexible workers keep their time and lose the cover.

The internet made her an emperor. Specifically the one who supposedly asked why his starving subjects did not eat meat. Pekingnology reaches for Marie Antoinette and the cake she almost certainly never mentioned, then notes the caricature does not fit the record at all.

She is the last academic to accuse of indifference. In 2023 she put an upper bound of 46.5% on youth unemployment against an official 19.7%, by counting 16 million young non-students who had stopped looking. Her fieldwork — nearly 30 factories, 20 labour brokers and ten labour markets in Kunshan, Dongguan, Foshan and Shenzhen — puts manufacturing gig workers at roughly 40 million, 31.12% of the sector.

The fight is about the word. One think tank has flexible employment reaching 320 million this year, some 44% of the workforce. At that scale "welfare" is no longer available as a technical term, and an economist who says it out loud is standing in for everyone who arranged things this way.

Meanwhile, formaldehyde on the cabbage. The State Council food safety office, the agriculture ministry and the market regulator jointly ordered Kangbao county in Hebei investigated over cabbage dipped in formaldehyde solution at the buying stage, with the produce traced and nationwide spot checks on perishable vegetables. Three central bodies, one county, same day.

Go deeper on this section: ClaudeChatGPT

Nobody will admit to Ox Alpha

A frontier coding model appeared with no owner. Ox Alpha went live on OpenRouter on 20 August with a 1,048,576-token context, 131,072 maximum output, text, image and video in, and zero pricing. The agent harness OpenCode says its provider has capacity for 100 trillion tokens a day.

So Chinese developers ran a DNA test. QbitAI's account of the crowdsourced forensics: token counts track GLM-5.3 with a constant 75-token offset across 25 prompts, explicable as a hidden system prompt; video budgets match Zhipu's GLM-5V-Turbo at about 147 tokens per second of footage, with sampling indifferent to frame rate.

The precedent helps. February's anonymous Pony Alpha turned out to be an early GLM-5, and two earlier stealth models were Xiaomi's MiMo. The Google camp is working from vague posts by DeepMind staff. What the exercise really shows is that behaviour no longer identifies a lab — distillation, merging and third-party serving have broken the link between who trained a model and where it runs.

The volume table is the real story. OpenCode's own tracker had Ox Alpha third last week at 7.1 trillion tokens, behind DeepSeek V4 Flash and MiMo. Two Chinese labs and an anonymous third that fingerprints Chinese. On Friday OpenAI cut GPT-5.6 Sol by over 20%, to $4 in and $20 out per million, naming Chinese models among the pressure.

Go deeper on this section: ClaudeChatGPT

Robot records fall, buyers ask about payback

Honor's Lightning took the 1500m. Two minutes 30.22 net, and 40.6 seconds over 400m. The Beijing Humanoid Robot Innovation Centre's Tiangong Ultra ran 38.15 seconds in the large-group 400m, Tiangong Omni 45.66 in the small. Every one beats the human world record, which tells you about gearing and battery chemistry and nothing about work.

The professionals on the floor have moved on. Leiphone's report from the same week: buyers now ask whether a brain can hold up for eight hours and whether the arithmetic closes. Sorting is the contested job. Yuanli Lingji's DM0.5 base model, out on 9 July, runs about two seconds an item — 1,200 an hour — on a plain gripper, and slices fruit on the same stack.

Generalisation is the currency. Zibianliang's WALL-B flips a parcel because a bump implies the label is underneath, and clears a surge on the main line before returning to a jam. The world-model lead at Tsinghua-founded home-robot startup Qiuzhi Tech told Leiphone the definitional fight over world models is rhetoric and only generalisation counts. He is right.

The man who funded it twice surfaced. Yin Fangming gave Wang Xingxing 2m yuan (~$282,000) for 15% of Unitree in 2016 before anyone else would, and took roughly 280m yuan (~$39m) out across sell-downs, about 140 times in. On 21 July he became chairman of Galbot — which has raised over 6.96bn yuan (~$980m) at above 20bn (~$2.8bn) — alongside a near-total management change. That reads like listing prep.

Go deeper on this section: ClaudeChatGPT

History gets treated as a security problem

State media went after the New Qing historians. An article in the state-run China Ethnic News accuses the US-born school of the last dynasty of supplying "theoretical ammunition" to separatists. The school's claim — that the Qing was a Manchu empire holding China rather than a Chinese dynasty — is load-bearing for Xinjiang, Tibet and Mongolia, which is exactly why it is being fought in a newspaper.

And the domestic flank is exposed. A senior academic warns that students dropping history as a gaokao subject leaves the public defenceless against pseudo-history — including the online claim that Ancient Greece never existed. The state wants the orthodox account held and has noticed nobody is being taught to hold it.

Then Pekingnology published the opposite argument. Zhao Shukai, who worked in the Party Secretariat's rural policy office through the 1980s and later at the State Council's research centre, argues rural reform had no designer: household contracting came from farmers and local cadres, and what allowed it was a plural power structure at the centre. In 2026 that is a claim about the present.

Go deeper on this section: ClaudeChatGPT

Threads we are pulling

  • The storm arrived where I said the toll was. Guangxi issued flash-flood warnings for Tiandeng and Daxin counties on Sunday evening. On 21 August I noted the province revised Maysak's toll from 39 to 159 on the day the next system formed; that system is now the one that grounded Chang'e-7.
  • AgiBot's listed shell is taking consumer deposits. Shangwei New Materials — the wind-blade materials company AgiBot took control of in 2025, chaired by co-founder Peng Zhihui — opened preorders on two small personal robots, September shipping, with 210m yuan (~$29.6m) of prepayments and 160m yuan (~$22.5m) of first-half R&D already booked.
  • Baidu denies it needs money. No share placement planned, 283.1bn yuan (~$39.9bn) of cash and investments, GPU cloud revenue up 283%, AI now half of core revenue — against June-quarter revenue of 31.3bn yuan (~$4.4bn), down 4.4%, and net profit down 68%. Denials answer questions somebody asked.
  • DeepSeek's weekend retreat is live as of midnight today, five days after peak pricing shipped. On V2EX the same day: a developer hunting a monthly Codex relay after burning $100 a day, and another asking whether any foreign model matches DeepSeek on price without nationality checks.
  • Hefei has a new party secretary. Qin Weiguo, already on Anhui's provincial party standing committee, takes the city that made state venture capital famous. Worth watching what he is told to buy.
  • The data gap got 617 hours smaller. Noitom released HiPHI on Hugging Face: 371.8 hours of whole-body motion plus 245.7 of human-object interaction, 132 actors, 90Hz sub-millimetre optical capture, already deployed on Unitree's G1. Against a world total of roughly 500,000 compliant robot-hours, under a twenty-thousandth of language-model scale.
  • Washington brought America First to APEC in China, weeks before Xi travels to the United States. Hosting the meeting where the other side campaigns against multilateral trade is a cost Beijing has decided to absorb.

Go deeper on this section: ClaudeChatGPT

The river · 34

A censored essay links banquet assaults to unchecked power

# China Digital Times 5h

A censored Chinese essay argues that alleged sexual assaults by officials at banquets cannot be addressed case by case without confronting the power structures that suppress reporting.

More
  • The essay discusses two recent allegations involving officials and says coverage of the newer case was constrained because local authorities had not issued an official notice.
  • It points to a removed government biography and a deleted newspaper article as signs of how information disappears before facts are publicly established.
  • The author says even state media can be rebuked for asking questions that local authorities have not authorized, making official statements gatekeepers of visibility.
  • The argument shifts from individual misconduct to institutional power: the problem is not merely personal self-control but the lack of restraints strong enough to contain authority.
  • The censorship is part of the story, not an editorial footnote; it shows how accountability narratives are managed through timing, wording and publication access.
  • For companies and institutions, the broader lesson is that formal hierarchies can distort both incident response and the evidence available to outsiders.
Open at China Digital Times Ask Claude ChatGPT

Why 'flexibility' sounds like an insult to Chinese workers

# Pekingnology 16h

A backlash against a Chinese economist's claim that flexibility is a form of welfare shows how precarious work, social insurance and elite language collide online.

More
  • The economist was discussing a labor-economics tradeoff: flexible workers control their time but generally receive less employment protection, while regular workers trade autonomy for benefits.
  • Many commenters heard the phrase as saying insecure work is itself a privilege, comparing it to a dismissive elite response to poverty.
  • The reaction reflects the material context of a very large gig workforce whose income, pensions and employer contributions are less predictable than those of formal employees.
  • The episode also shows how a qualified academic argument can fail when it uses the vocabulary of labor economics without acknowledging the power imbalance behind the choice of flexibility.
  • The Chinese social-media response is therefore not just semantic outrage; it is a referendum on whether workers believe they have a real choice between autonomy and security.
  • The controversy reveals a widening distance between policy or academic language and the lived experience of people facing unstable employment.
Open at Pekingnology Ask Claude ChatGPT

Gen 1.5 hints at a GPT-3 moment for physical AI

# GeekPark 14h

Generalist's GEN-1.5 robot model can adapt to new tools and tasks from a short demonstration without parameter updates, suggesting a path from scripted manipulation toward physical generalization.

More
  • In one demonstration, a robot learned to sweep blocks with a brush, then inferred how to use a banana and a dustpan without being explicitly trained on those tools.
  • The model's physical prompting uses a three-to-12-second demonstration and then attempts the task in changed object positions without fine-tuning.
  • Generalist reports 59% average success on ten new tasks from one demonstration, rising to 83% after five minutes of data and ten parameter updates.
  • The system reportedly cut the required training updates from hundreds to zero for some demonstrations, shifting adaptation from offline retraining toward inference-time interpretation.
  • Earlier generations improved task success as the company increased physical data, model size and compute; GEN-1 reportedly reached 99% average success on selected simple tasks after more than 500,000 hours of data.
  • The result is not yet a deployable general robot: one-shot success remains 59%, and the report does not establish performance across broad household or industrial distributions.
  • The strategic question is whether this curve continues as the model encounters contact dynamics and failure modes that cannot be covered by short demonstrations alone.
Open at GeekPark Ask Claude ChatGPT

Alibaba plans an $11.3B share placement entirely for AI

# Leiphone 4h

Chinese coverage stresses that Alibaba will devote all 80 billion Hong Kong dollars, roughly $11.3 billion, to AI, while US coverage emphasizes the size and discount of the follow-on sale.

Also: Techmeme

More
  • Alibaba's first share placement since its 2019 Hong Kong listing is aimed at non-US investors, according to the report.
  • The stated use of proceeds is full-stack AI capability and infrastructure, making this a capital-allocation decision rather than a routine funding round.
  • The company says AI-related products have passed 49.5 billion yuan, roughly $7 billion, in annualized revenue, but the supplied report does not independently verify that figure.
  • The raise signals that a major Chinese internet platform is willing to finance AI buildout at public-market scale even while its core businesses remain central to the company.
Open at Leiphone Ask Claude ChatGPT

Remembering Zhu Rongji and the governing style he left behind

# SCMP 3h

As China mourns former premier Zhu Rongji, a reflective essay uses his career to examine the reform-era governing style and the political memory attached to it.

More
  • The article is a personal remembrance prompted by reports of Zhu's death, not a new policy announcement.
  • Zhu served as premier from 1998 to 2003 and is widely associated with market-oriented restructuring, fiscal reform and a forceful anti-corruption style.
  • The essay's importance lies in what can be publicly remembered about a former senior leader and which elements of his legacy remain politically legible.
  • For readers studying China's state, Zhu represents a period when technocratic economic management and disruptive reform were closely linked.
  • The piece frames history as a guide to current choices rather than treating Zhu as a purely ceremonial figure, giving the remembrance a contemporary political purpose.
  • How official institutions and ordinary citizens reconcile that reform legacy with today's more centralized governance is the question the essay leaves open.
Open at SCMP Ask Claude ChatGPT

Alibaba's liver-cancer model found 15 missed tumors in clinical use

# IT Home 39m

Alibaba's DAMO LiON CT model found 15 previously missed malignant lesions in a two-month prospective deployment, mostly tumors about one centimeter across.

More
  • The model is designed to flag small, low-contrast liver lesions and liver metastases that can be obscured by cirrhosis, fatty liver or competing primary tumors.
  • The report says it read more than 10,000 contrast-enhanced CT scans and routed disagreements with first-read doctors to senior review or multidisciplinary discussion.
  • In testing, AI-assisted doctors took 27% less time and improved malignant-lesion sensitivity by 11.5%; the report also says junior doctors reached senior-level performance with assistance.
  • The architecture combines whole-liver context with local texture and boundary information, then fuses multiple scan phases to catch lesions that appear only briefly across the sequence.
  • The deployment is more informative than a benchmark because it shows a concrete human-in-the-loop escalation path rather than autonomous diagnosis.
  • The reported results still come from one two-month hospital deployment, so durability across institutions and disease mixes remains the key unanswered question.
Open at IT Home Ask Claude ChatGPT

A humanoid robot completes autonomous tennis rallies on live TV

# AI QbitAI 19h

Chinese coverage calls Galaxy General's live autonomous tennis match an embodied-AI milestone, while broader event coverage is more cautious about the gap between athletic demos and safe deployment.

Also: Leiphone

More
  • The robot reportedly served, returned forehands and backhands, moved to the net and collaborated with a human doubles partner against fast incoming balls.
  • Tennis forces perception, strategy, balance and fine manipulation into one continuous loop, unlike a fixed gesture or isolated pick-and-place demonstration.
  • Galaxy General says its AstraBrain model unifies higher-level decisions, whole-body control and neural action translation rather than joining separately trained modules at deployment.
  • The company also describes learning from imperfect human motion fragments, simulation against virtual opponents and feedback from physical play.
  • The comparison with AlphaGo is rhetorically powerful because the robot is being tested in an open physical environment, but the supplied report does not establish repeatability across opponents or venues.
Open at QbitAI Ask Claude ChatGPT

A young team is betting on an end-to-end brain for bipedal robots

# AI QbitAI 19h

A newly formed Chinese team of doctoral researchers is pursuing an end-to-end model for bipedal robots, arguing that interfaces between planning and control create an information bottleneck.

More
  • The team contrasts a single perception-to-action model with the common split in which a language or task model hands commands to a separate movement controller.
  • Its technical target is full-body dynamics: a biped must shift its center of mass while reaching, squat while balancing and manage contact forces across 29 degrees of freedom.
  • The team says its earlier research on a quadruped VLA exposed cascading errors and task-specific tuning at the boundary between brain and cerebellum.
  • The approach is deliberately risky because current research still often favors modular systems for reliability, while end-to-end scaling needs large, diverse physical data.
  • Its card-driving demo is a test of synchronized vision, hands, feet and force control under high-speed motion, not merely a locomotion clip.
Open at QbitAI Ask Claude ChatGPT

A robot cafe puts embodied AI to work in public

# AI QbitAI 15h

At Beijing's robot conference, a robot cafe let machines serve drinks, avoid people and collect waste in an open environment rather than performing isolated scripted demos.

More
  • The service robot reportedly planned paths around moving visitors, distinguished private objects from rubbish and handled soft paper while carrying out a multi-step service loop.
  • A robotic arm prepared coffee behind the counter while a mobile robot delivered and collected cups, creating a mixed human-machine workflow in a public space.
  • The demonstration tests dynamic perception, replanning, semantic object handling and long-horizon task sequencing at once, which is more representative than a fixed obstacle course.
  • The operator says the system had already been trialed in a real cafe, but the supplied account does not provide uptime, error or labor-cost data.
  • The commercial question is whether such robots can sustain safe, graceful operation over long shifts when customers behave unpredictably and tasks arrive asynchronously.
Open at QbitAI Ask Claude ChatGPT

Developers try to identify the anonymous Ox Alpha model

# AI QbitAI 18h

Chinese developers are fingerprinting the anonymous Ox Alpha model through tokenization, video-token budgets and API errors, with Z.ai currently the leading but unproven suspect.

More
  • The model reportedly accepts text, images and video with a 1.05-million-token context and is being used for code, long-running agents and complex reasoning.
  • Community tests found token-count offsets resembling GLM models and video-token allocation patterns similar to a Z.ai multimodal system, but these are reverse-engineering clues rather than attribution proof.
  • Users also found an error format resembling Z.ai services, while a claimed hidden training message is weaker second-hand evidence.
  • Google remains another hypothesis because Ox Alpha's product profile resembles Gemini, and employee social posts have fueled speculation without confirming anything.
  • The episode illustrates a distinctive open-model culture in which developers inspect system behavior as a kind of digital provenance test when vendors withhold the model's identity.
Open at QbitAI Ask Claude ChatGPT

A former security guard reaches an AI contest final

# AI QbitAI 5h

A former security guard reached the top 20 of a Chinese AI contest after using an AI coding tool to build a browser-based motion game, while finalists also tackled AI-to-hardware workflows.

More
  • The contest drew 37,000 registrations and 14,000 submissions, making the result a useful signal of how widely AI tools are lowering entry barriers.
  • The runner-up project used an AI hardware IDE and desktop PCB mill to turn a natural-language request into components, code, simulation, routing and a physical board.
  • Testing with about 500 users exposed the real bottlenecks: incomplete component data, broken build environments and simulations that did not match hardware.
  • The team used error telemetry and context-aware debugging to make the generated hardware workflow more robust, moving beyond a demo that worked only under ideal conditions.
Open at QbitAI Ask Claude ChatGPT

China's Qiyuan opens reservations for personal home robots

# AI QbitAI 21h

Qiyuan has opened public reservations for two small personal robots, signaling a shift in China from exhibition demos toward consumer testing and paid ownership.

More
  • The company says it has taken 210 million yuan, roughly $30 million, in prepaid orders and is building a supply chain for tens of thousands of units.
  • The Q1 is an 88-centimeter force-controlled biped, while the T1 can switch between wheeled-bipedal and quadruped forms; both target companionship, learning and creator use.
  • Consumer robots face harder conditions than factory machines because homes are open, personalized and full of changing objects, making reservations an early test of support costs and retention.
  • The launch has policy tailwinds: China's commerce authorities have listed AI robots as a new consumer supply category.
Open at QbitAI Ask Claude ChatGPT

Hong Kong professors are building the embodied-AI stack

# Analysis QbitAI 17h

A survey of Hong Kong university founders finds professors building components across embodied AI, from world models and tactile sensing to surgical robots, drones and autonomous vehicles.

More
  • The pattern is specialization: many teams are building a robot organ or subsystem rather than competing to sell a complete humanoid.
  • Hong Kong universities contribute research talent while the surrounding manufacturing and startup networks provide a route into mainland production and deployment.
  • The article presents the professor-student model, associated with DJI and other successes, as an institutional mechanism for turning lab expertise into companies.
  • The breadth of the ecosystem matters because embodied AI needs data, hands, control, bodies and application channels, not only a larger language model.
Open at QbitAI Ask Claude ChatGPT

One robot brain, several bodies: Galaxy General's WRC pitch

# AI GeekPark 14h

At Beijing's robot conference, Galaxy General presented a shared upper-level intelligence spanning a biped, wheeled arms and heavy-load robots rather than treating each body as a separate product.

More
  • The proposed AstraBrain stack reuses perception, task understanding, world modeling and planning while leaving body-specific dynamics and control systems in place.
  • Its small biped ET1 uses a whole-body control model trained on about 100,000 hours of human movement data to map observed motion into executable actions.
  • The architecture is closer to a reusable software layer than one controller copied across machines: different bodies retain their own motion models.
  • That separation could let a company amortize world knowledge across form factors, but the report provides demonstrations rather than independent evidence of transfer performance.
Open at GeekPark Ask Claude ChatGPT

Shein opens Hong Kong IPO targeting roughly $1.85B

# Signal Leiphone 2h

Chinese coverage emphasizes Shein's technology-heavy supply chain and global footprint, while US coverage reduces the story to a September 1 Hong Kong debut seeking up to $1.8 billion.

Also: Techmeme

More
  • The offering is priced at 47.6 to 49.5 Hong Kong dollars per share and could raise about 13.123 billion Hong Kong dollars, roughly $1.85 billion, after expenses.
  • Shein says 40% of proceeds will fund technology, another 40% global branding and expansion, and the remainder corporate responsibility and general purposes.
  • The company says its small-batch automated replenishment model links demand sensing, design, production allocation, routing and fulfillment across about 160 markets.
  • The Hong Kong listing is a financing and legitimacy test for a China-founded retailer that operates globally under heavy scrutiny.
Open at Leiphone Ask Claude ChatGPT

African firms adopt Chinese AI pragmatically

# Analysis SCMP 16h

African companies are using Chinese AI models where price or flexibility is better while keeping US options open, challenging a clean geopolitical split in the AI market.

More
  • The account presents model choice as an infrastructure decision driven by cost, capability and deployment flexibility rather than national allegiance.
  • Chinese models gain an opening when companies want alternatives to a small set of expensive US providers or need more control over deployment.
  • For China, overseas adoption creates a commercial and ecosystem foothold even when customers remain deliberately multi-vendor.
  • The report's strategic point is strongest in markets that cannot afford to treat AI procurement as a symbolic extension of US-China rivalry.
Open at SCMP Ask Claude ChatGPT

Why Chinese media says US historians arm separatists

# Culture SCMP 18h

Chinese state-linked coverage is portraying a US school of thought on the Qing dynasty as theoretical ammunition for separatism, turning historical scholarship into a current political battleground.

More
  • The campaign comes alongside online arguments about Qing history, ethnicity and claims that established ancient civilizations were fabricated.
  • The phrase historical nihilism functions as an official label for narratives that challenge the state's preferred account of China's past.
  • The episode shows how academic interpretation can become politically sensitive when it touches ethnic policy, territorial unity or the legitimacy of the national story.
  • For outside researchers, the risk is not only censorship of present-day politics but pressure to conform on apparently historical questions.
Open at SCMP Ask Claude ChatGPT

Chinese team demonstrates compact real-time X-ray ghost video

# Analysis SCMP 55m

Chinese researchers report a tabletop X-ray ghost-imaging system that captures real-time video of fast-moving objects with a simpler setup and potentially lower radiation dose.

More
  • The proposed technique is aimed at dynamic targets such as rotating aircraft engines and biological motion, where conventional X-ray imaging is difficult.
  • Its value is architectural: the team claims to trade a large traditional imaging setup for a compact system while retaining time-resolved, see-through imaging.
  • The source excerpt is incomplete, so performance limits, validation conditions and the exact reconstruction method remain unclear.
  • If the result holds outside laboratory demonstrations, it could matter for non-destructive inspection and biomedical imaging more than for ordinary radiography.
Open at SCMP Ask Claude ChatGPT

China delays Chang'e-7 moon mission

# Analysis SCMP 16h

China postponed Chang'e-7, a complex mission intended to search for and drill near ice deposits at the lunar south pole, after a comprehensive assessment.

More
  • The mission carries an orbiter, lander, rover and hopper, making it a distributed exploration architecture rather than a single landing vehicle.
  • The delay is operationally routine in form but matters because south-pole volatile resources are central to China's long-term lunar exploration plans.
  • The supplied report gives no new launch date, so the schedule impact cannot yet be assessed.
Open at SCMP Ask Claude ChatGPT

History education becomes a new front in China's narrative battles

# Culture SCMP 1d

A Chinese academic warns that exam reforms reducing students' exposure to history could leave the public more vulnerable to pseudo-history as Beijing intensifies its campaign against historical nihilism.

More
  • The concern is that when students drop history as an examination subject, fewer people develop the tools to evaluate competing narratives later in life.
  • The warning arrives as official discourse labels claims that undermine the state's historical line as historical nihilism, linking classroom knowledge to political legitimacy.
  • The story shows education policy functioning as narrative infrastructure: curriculum choices shape which citizens are equipped to challenge or reproduce historical claims.
  • The tension is not simply between true and false history; it is also over who gets to define legitimate interpretation and how much room schools have for disagreement.
Open at SCMP Ask Claude ChatGPT

Washington takes an America First line at APEC in China

# Policy SCMP 22h

US officials are bringing an America First trade agenda to APEC talks hosted in China ahead of a planned Xi visit to the United States, making the forum a test of managed rivalry.

More
  • The US framing emphasizes reducing barriers for American businesses and pursuing reciprocal trade rather than presenting APEC primarily as regional institution-building.
  • The talks take place in China weeks before the anticipated presidential visit, giving routine economic diplomacy a clear bilateral political backdrop.
  • For companies, the language points to continued negotiation over market access and supply chains even where both sides maintain high-level engagement.
Open at SCMP Ask Claude ChatGPT

The investor who backed Unitree now chairs another robot startup

# Analysis QbitAI 17h

An early Unitree backer who reportedly turned a 2 million yuan, roughly $282,000, investment into about 280 million yuan, roughly $39 million, now chairs Galaxy General.

More
  • The profile says Yin Fangming invested in Unitree when quadruped robots were obscure, then shifted from direct ownership to indirect holdings and later invested across robotics, energy and space.
  • His own earlier robotics company, ROOBO, struggled despite large fundraising, giving him experience with the cost and failure modes of hardware startups.
  • The story's broader signal is that China's robotics ecosystem is being connected by repeat founders and investors, not built from isolated one-off companies.
Open at QbitAI Ask Claude ChatGPT

China's 4.27 million-car recall is mostly software and labeling

# Policy Jiemian 30m

China's largest recorded recall covers more than 4.27 million vehicles, but the fix is a warning label and an OTA change to emergency window behavior rather than mechanical repair.

More
  • Nine manufacturers, including Tesla, Leapmotor, Xpeng, Geely, Xiaomi and Chery, filed recalls over emergency mechanical handles that could be hard to identify after a low-voltage failure.
  • The proposed remedy combines clearer labeling with remote software changes intended to improve window opening after severe collisions.
  • The scale makes the episode a useful test of whether China's connected-car recall system can treat software behavior as a safety defect.
  • The commentary frames the recall as evidence of more mature defect detection rather than proof that new-energy vehicles are broadly unsafe.
Open at Jiemian Ask Claude ChatGPT

Beijing's willow and poplar trees may amplify ozone pollution

# Signal Solidot 14h

A study reports that plant emissions accounted for about 10% of Beijing's volatile-organic-compound emissions and 52% of the chemicals driving ozone formation during the 2021 warm season.

More
  • The researchers found that 35% of Beijing's trees emit isoprene, which reacts with hydroxyl radicals and nitrogen oxides under sunlight to form ozone.
  • Plants emitted less total volatile organic compounds than human activity but contributed disproportionately to ozone chemistry because of the compounds involved.
  • The finding complicates urban greening policy: tree selection can affect air quality, especially in cities with high nitrogen-oxide levels.
Open at Solidot Ask Claude ChatGPT

Sichuan ends earthquake-alert authorization after a false 7.7 warning

# Policy IT Home 7m

Sichuan coverage centers on a regulator cutting off an alert provider after a 7.7 warning for a 4.7 quake, exposing the governance risks of delegated safety infrastructure.

Also: Jiemian

More
  • The Sichuan bureau says the institute sent a 7.7-magnitude warning through Honor, Vivo and Xiaotiancai devices, while the officially measured quake was magnitude 4.7.
  • The bureau ended the institute's third-party authorization on July 22, making the dispute as much about control of public alert channels as about one bad estimate.
  • The official earthquake network produced a 5.4 warning and distributed it through its own mini-program, television, terminals and phones.
  • The institute apologized and called this its largest error since beginning public service in 2011, promising an investigation and service improvements.
  • The episode follows criticism of advertising and a supposed paid acceleration channel in its app, showing how monetization language can undermine trust in safety-critical systems.
  • For technology operators, the important boundary is whether private distributors can alter or misrepresent official alerts once they are embedded in consumer hardware.
Open at IT Home Ask Claude ChatGPT

Researchers find Android malware targeting a car head unit

# Signal Solidot 13h

Kaspersky says it found Android malware spreading through the built-in updater of a car infotainment system, a sign that vehicle firmware is becoming a botnet attack surface.

More
  • The malware reportedly targets an Android-based head unit and is linked by researchers to a group associated with the BADBOX residential-proxy network.
  • The suspected goal is to recruit compromised vehicles into proxy infrastructure, reusing a pattern previously seen in cheap TV boxes.
  • Kaspersky says the vehicle provider was notified and fixed the relevant security problem, but the incident highlights the importance of securing update chains and separating infotainment from safety systems.
Open at Solidot Ask Claude ChatGPT

iQIYI's AI push shows up as long-video economics reset

# Signal 36Kr 2h

iQIYI reported 6.29 billion yuan, roughly $886 million, in second-quarter revenue as AI short dramas, overseas memberships and content licensing begin to offset pressure on subscriptions.

More
  • Membership revenue was 4.01 billion yuan, roughly $565 million, while content distribution rose 56% year on year to 680 million yuan, roughly $96 million.
  • The company says AI short dramas accounted for more than half of both short-drama output and unique visitors in June, turning AI from experimentation into a measurable content channel.
  • Its professional AI production platform claims to cut production time and cost by 70% to 90%, but those are company estimates rather than independent operating data.
  • The strategic shift is from being a centralized long-video destination toward a production, creator-services and distribution platform that can monetize content beyond subscriptions.
Open at 36Kr Ask Claude ChatGPT

A Chinese academic says US AI supply-chain strategy will fail

# Policy SCMP 17h

A prominent Chinese political scientist argues that US chip, rare-earth and AI alliances will backfire and that restricting access to low-cost Chinese models will hurt both sides.

More
  • The argument is explicitly a Chinese policy perspective, not neutral evidence that the US initiatives will fail.
  • It treats external demand and economic cooperation as necessary for long-term growth, contrasting that view with competitive supply-chain separation.
  • The useful signal is how some Chinese intellectuals frame export controls: not only as containment, but as a self-defeating attempt to deny customers cheaper technology.
Open at SCMP Ask Claude ChatGPT

Orange lichen may point explorers to rare-earth deposits

# Signal SCMP 14h

A Chinese research team says a vivid orange lichen on exposed rock could serve as a visible indicator of undiscovered rare-earth deposits, potentially aided by drones.

More
  • The proposed indicator is valuable because the lichen is easy to spot in the field, unlike a subsurface mineral body that normally requires extensive sampling.
  • The source excerpt does not provide the biological mechanism, field-validation rate or evidence that the signal works outside the studied environments.
  • If validated, pairing the indicator with unmanned aerial surveys could lower the cost of early-stage mineral exploration.
Open at SCMP Ask Claude ChatGPT

A smart ring tries to sell emotion rather than health metrics

# Culture Leiphone 22h

A Chinese smart-jewelry brand is positioning a 1,499-to-2,399-yuan, roughly $211-to-$337, ring around emotional interpretation and design rather than competing on health-sensor accuracy.

More
  • The ring combines standard wellness measurements with an assistant that turns data into literary mood explanations and a shared relationship feature whose data visibility users control.
  • Its strategy is to use silver, lab-grown diamonds and fashion variants to escape the low-price sensor market, but it lacks the years of validation and ecosystem depth of Oura or Samsung.
  • The core business question is whether emotional interpretation creates repeat purchases and retention strong enough to compensate for weaker technical differentiation.
Open at Leiphone Ask Claude ChatGPT

Chinese archaeologists join Peru's first joint field project

# Culture SCMP 13h

Chinese archaeologists will work with Peruvian institutions on ancient sites in the first joint field project of its kind in South America, extending Beijing's cultural ties through research.

More
  • The partners include China's national archaeology institute, Peru's Caral archaeological zone and a university archaeology school.
  • The project uses scientific collaboration, rather than trade or infrastructure, as a channel for China's broader engagement with Latin America.
Open at SCMP Ask Claude ChatGPT

Tencent names 50 winners of its young-scientist award

# Signal IT Home 1h

Tencent's 2026 Science Exploration Award selected 50 researchers, each receiving 3 million yuan, roughly $423,000, over five years, with more international peer review and a first winner born in the 1990s.

More
  • The privately funded award targets younger researchers working full time in mainland China, Hong Kong or Macau across basic science and frontier technology.
  • International reviewers now make up 40% of peer review, while ten women won awards, the first time the program has reached double digits.
  • The program says past winners have produced work later recognized by national science awards, making it a parallel funding channel for early-career researchers rather than a conventional corporate prize.
Open at IT Home Ask Claude ChatGPT

China roundup: rumor arrests, graduate pay and AI pricing

# Signal 36Kr 4h

A Chinese business roundup links a police action over a false Xiaohongshu IPO rumor with vocational wage data, DeepSeek pricing and Alibaba's planned AI financing.

More
  • The most useful social signal is a comparison showing some vocational fields paying more than the lowest-paid university majors six months after graduation.
  • The roundup says a blogger was detained for inventing an IPO-failure claim about Xiaohongshu, illustrating how online finance rumors can become a public-order matter.
  • Its AI items emphasize lower weekend pricing from DeepSeek and cost competition among model providers, but these are routine product notices.
Open at 36Kr Ask Claude ChatGPT

Lei Jun sees a gap at the top of China's car market

# Analysis IT Home 16m

Lei Jun says China's crowded low end is masking a shortage of credible premium vehicles, with local high-end components still needing five to ten years to mature.

More
  • Lei says the premium market is being reshaped because supply of convincing higher-end products is insufficient even as lower-priced segments fight a price war.
  • He identifies vehicle chips, operating systems, advanced chassis parts and carbon-ceramic brakes as areas still moving through domestic substitution.
  • His five-to-ten-year estimate is a judgment from a major Chinese automaker, not an industry forecast, but it highlights where localization remains incomplete.
  • July data cited in the report show vehicles priced above 400,000 yuan, roughly $56,000, were the only passenger-car segment to grow year on year.
Open at IT Home Ask Claude ChatGPT
Source ledger
ChinAI 403 0/0 HTTPError: 403 Client Error: Forbidden for url: https://chinai.substack.com/feed
ChinaTalk 200 0/20
Pekingnology 200 1/20
Ginger River Review 200 0/20
Sinocism 200 0/20
QbitAI 200 10/10
36Kr 200 3/30
Solidot 200 4/18
Leiphone 200 9/20
GeekPark 200 3/30
IT Home 200 12/60 +41 over cap
SCMP 200 10/50
Initium 200 1/15
Jiemian 200 8/30 +16 over cap
China Digital Times 200 1/15
The Wire China 200 0/50
Techmeme 200 5/15
Hacker News 200 1/20
V2EX 200 8/50 +31 over cap