China Desk — 20 Aug 2026, 01:46 UTC

Unitree's founder outlines a self-improving robot training loop

# GeekPark 13h

After Unitree's listing, founder Wang Xingxing described a system linking model search, code generation, simulation, real-robot tests, and evaluation into a reusable learning loop.

The proposed loop is search for papers and code, generate training programs, simulate them, deploy to physical robots, score the results, and feed the evidence into the next iteration.
This is an architecture for an engineering organization as much as a model: the AI participates in research, training, deployment, and evaluation rather than serving only as the robot's controller.
The hard bottleneck is physical feedback. Simulation can generate cheap rollouts, but balance, contact, failure, and task success must be measured on real hardware.
Unitree has sold more than 30,000 quadrupeds and shipped more than 5,500 humanoids in 2025, giving it a larger installed manufacturing base than many prototype-stage rivals.
The company controls motors, joints, motion control, and system integration, allowing hardware and policy changes to be co-designed instead of negotiated across suppliers.
A major unresolved issue is data rights: sold robots do not automatically provide usable training data, so consent, standardization, and secure feedback pipelines must be solved.
If Unitree can deploy thousands of robots and recycle validated skills across them, its manufacturing scale could become an AI data advantage rather than merely a hardware advantage.

Open at GeekPark Ask Claude ChatGPT

Security checks at Beijing's Qianmen landmark trigger censorship

# China Digital Times 20h

A deleted Chinese travel post describing intense security checks at Qianmen was blocked under cybersecurity rules, illustrating censorship's collateral reach.

The post reportedly described identity scans, bag and book checks, pat-downs, and other visitor screening at a landmark beside Tiananmen Square.
It contained no overt political criticism; its apparent offense was documenting an ordinary encounter with a sensitive security regime.
The episode shows how broad rule enforcement can suppress mundane logistical accounts when automated systems or complaint-driven review associate them with politically sensitive places.
For visitors and platforms, the practical uncertainty is not only what content is prohibited but whether neutral description can be treated as a security violation.
The source is a translation of deleted material, so the exact trigger for removal is unclear; it explicitly notes that automated keyword scanning may have contributed.

Open at China Digital Times Ask Claude ChatGPT

China successfully recovers the first stage of a reusable rocket

# Solidot 18h

LandSpace's Zhuque-3 placed a satellite into orbit and softly landed its first stage, China's first land recovery of an orbital-class rocket stage.

The rocket's first stage landed six minutes after launch using nine engines, in a profile broadly comparable to Falcon 9's propulsive return architecture.
The event gives China a flight-validated reusable launch vehicle alongside Falcon 9, New Glenn, and the reported Changzheng 10B system.
The important distinction is land recovery: prior Chinese validation cited in the source used a sea-based net, while this mission used landing legs at a prepared inland site.
The technical milestone is not yet a reusable service. Turnaround time, refurbishment cost, payload penalty, and repeated flights will determine whether the architecture changes launch economics.
The mission carried Honghu 03 into orbit, showing that recovery was integrated with an operational orbital flight rather than tested as an isolated hop.

Open at Solidot Ask Claude ChatGPT

Beijing bars Chinese firms from assisting the EU's JD.com probe

# SCMP 12h

Beijing ordered Chinese organizations and individuals not to assist an EU foreign-subsidy investigation into JD.com's proposed acquisition, escalating a regulatory clash.

China's Ministry of Justice called the EU investigation an improper exercise of extraterritorial jurisdiction and prohibited assistance with the probe.
The order creates a direct conflict of legal obligations for companies, auditors, advisers, and data holders operating across both jurisdictions.
The move signals that Beijing is willing to use domestic law and administrative authority to contest foreign scrutiny of Chinese companies, rather than treating the investigation as a private transaction matter.
For cross-border acquisitions, diligence access and information rights may become geopolitical risks alongside antitrust and national-security review.
The source says the EU opened an in-depth investigation in May, but the excerpt does not provide its findings or the acquisition's final status.
The practical test is whether Brussels narrows the probe, seeks alternative evidence, or treats Beijing's order as grounds for stronger enforcement.

Open at SCMP Ask Claude ChatGPT

Chinese researchers entangle quantum memories across 420 kilometers

# Solidot 15h

A University of Science and Technology of China team demonstrated entanglement between cold-atom quantum memories over 420 kilometers of fiber, beyond the direct-distribution limit.

The result connects two matter-based quantum memories across a fiber distance that exceeds the reported 230-kilometer theoretical limit for unrepeated entanglement distribution.
Quantum memories are important because future repeaters must store and synchronize fragile quantum states rather than simply transmit classical signals.
The work builds on a progression from 50 kilometers in 2020 to a three-node metropolitan network in 2024 and higher-fidelity long-distance demonstrations in 2026.
Potential applications include repeater-based quantum communications, distributed quantum computing, and distributed sensing, but none is yet a commercial service.
The engineering question is whether fidelity, synchronization, loss tolerance, and component reliability can scale from a laboratory link to a network with many repeaters.

Open at Solidot Ask Claude ChatGPT

China completes first beam test for a medical isotope accelerator

# IT Home 6m

A Chinese superconducting linear accelerator reached a 93 MeV, 1 mA beam, advancing domestic mass production of cancer-treatment isotopes.

The IP-SAFE facility in Lanzhou is designed to produce actinium-225 and radium-223, both alpha emitters used in targeted radiopharmaceutical research.
The initial beam test validated the accelerator and supporting process systems; the source says the facility now has the basic capability for batch production, not that full commercial output has begun.
The team combines thorium-target irradiation, automated separation and purification, and drug-labeling processes in one production route.
The strategic objective is supply-chain autonomy in isotopes whose short half-lives and difficult production make imported supply vulnerable.
The next test is integration across the accelerator, isotope target, and separation systems; throughput, purity, regulatory approval, and clinical supply reliability will determine the real impact.

Open at IT Home Ask Claude ChatGPT

Jefferies test ranks Alibaba's Qwen Office first among eight agents

# AI QbitAI 39m

A Jefferies office-task test ranked Qwen Office first, highlighting harness engineering and cost per completed task rather than model scores alone.

Also: Leiphone

The five tasks covered document synthesis, web research, browser control, English presentation creation, and image-based marketing design.
The report says Qwen Office was the only system to score above 90 across every tested dimension, though the methodology and full score table are not included in the source.
Its harness score exceeded those of Claude Cowork and Codex, reinforcing that tool routing, context handling, correction loops, and governance can dominate raw model capability.
The underlying Qwen 3.8 Max API is described as cheaper than some leading foreign models, making execution cost important for long, tool-heavy enterprise workflows.
Qwen Office has begun connecting to DingTalk for chat summaries, documents, spreadsheets, email, and other workplace actions.
The result is a useful China-specific signal, but one bank's benchmark should not be treated as a general ranking of agent intelligence.

A second test also puts Qwen Office first among global agents Leiphone

Open at QbitAI Ask Claude ChatGPT

Zhu Rongji's four rules for state-run investigative journalism

# Culture Pekingnology 7h

A translated 1998 speech by former premier Zhu Rongji shows how China's top leadership once used investigative television to expose local corruption while keeping media within Party authority.

Zhu praised the state broadcaster's investigative program for exposing bureaucratic abuse and said officials should expect public scrutiny from it.
The speech reveals a controlled accountability model: media can attack local wrongdoing when that serves central legitimacy, but the institution remains part of the state system.
His emphasis on evidence, persistence, and public support offers a historical contrast to today's more security-centered information environment.
Publishing the old speech now is itself informative, because historical examples of supervision and criticism are often revisited selectively in debates about governance.
The source is a translation and contextual essay, not evidence that the 1998 media model has returned.

Open at Pekingnology Ask Claude ChatGPT

Perfect World says new game has passed $280 million in global revenue

# Signal QbitAI 25m

Perfect World's new game exceeded 2 billion yuan (roughly $282 million) in global player spending, even as launch marketing pushed the company to a first-half loss.

Perfect World reported first-half revenue of 2.751 billion yuan (roughly $388 million), a net loss of 118 million yuan (roughly $16.6 million), and operating cash flow of 647 million yuan (roughly $91 million).
The company attributes the loss partly to marketing costs being recognized immediately while player spending is recognized over the game's lifecycle.
The game launched across more than 180 countries and six device categories, with official PC and Android channels representing nearly 60% of reported spending.
The result illustrates how Chinese game publishers are using simultaneous global launches and multi-platform distribution to reduce dependence on the domestic market.
The important uncertainty is retention: strong launch spending can become durable franchise economics only if content updates keep players active after acquisition costs normalize.

Open at QbitAI Ask Claude ChatGPT

Chinese homebuyers who tried to call the bottom kept waiting

# Analysis Initium 2h

Chinese buyers who bought during the property downturn have not found a clear market bottom, exposing the limits of timing a prolonged housing correction.

The story centers on households that treated falling prices as a temporary opportunity and purchased before the decline had stabilized.
The broader signal is behavioral: repeated expectations of a quick rebound can keep buyers, developers, and local governments misreading the depth of the adjustment.
China's property market is not only an asset story; it affects household balance sheets, local fiscal revenue, construction employment, and confidence.
The article's reporting is narrative rather than a new price dataset, so it illustrates sentiment more clearly than it measures national conditions.
For technology companies, weak housing confidence matters through consumer spending, collateral availability, and local-government willingness to fund projects.

Open at Initium Ask Claude ChatGPT

Chinese space-compute startups confront a four-week upload problem

# Signal GeekPark 21h

Chinese space-compute companies are pursuing orbital AI, but sending an 800 MB model to an in-orbit satellite can still take nearly four weeks over low-rate uplinks.

The bottleneck is not simply onboard compute: traditional command links operate around 1 Mbps and offer only short ground-station windows on each orbit.
Downlink has improved because ground receivers can correct atmospheric distortion, while high-speed uplink remains difficult and lacks an established operational architecture.
The report distinguishes processing data already generated in orbit from sending ground data upward for training; only the latter creates the much larger infrastructure challenge.
More than 30 Chinese companies are reportedly exploring the sector, but the current engineering reality makes rapid model iteration and software updates impractical.
The immediate commercial wedge may therefore be 'space data storage' or edge filtering rather than a general orbital data center.
Solving uplink throughput, optical link availability, thermal management, radiation tolerance, and launch economics will determine whether the sector becomes infrastructure or remains a narrative.

Open at GeekPark Ask Claude ChatGPT

Why Chinese audiences are flocking to a film they dislike

# Culture SCMP 16m

A Chinese animated film jumped from 3,420 yuan to 11 million yuan in box-office revenue, showing how online curiosity can suddenly reshape cinema demand.

The film opened to almost no commercial traction, then experienced a sharp audience surge roughly ten days later and entered the top tier of domestic animated releases.
The pattern is less a conventional hit than a case study in virality: attention, discussion, and the pleasure of watching an unlikely failure can become the product.
Chinese cinema increasingly depends on platform-driven discovery, where a meme or controversy can overcome weak initial distribution and marketing.
The available report does not establish whether the spike reflects sustained word of mouth, organized viewing, or a short-lived online phenomenon.
The useful signal for media and consumer businesses is that demand can be created after release, but its conversion from attention to durable revenue remains uncertain.

Open at SCMP Ask Claude ChatGPT

Wang Yi urges Washington to change its North Korea policy

# Policy SCMP 33m

Chinese coverage presents Wang Yi's Seoul talks as a call for peaceful coexistence and a more constructive US approach to Pyongyang.

China's foreign minister said Beijing supports South and North Korea moving toward peaceful coexistence and portrayed China as a stabilizing actor on the peninsula.
The framing places the burden on Washington, describing its policy toward Pyongyang as hostile rather than focusing on North Korean conduct.
Wang's meeting with South Korea's foreign minister also serves a bilateral purpose: Beijing is signaling that Seoul should maintain engagement with both neighboring powers.
The timing, ahead of an anniversary of China-South Korea diplomatic ties, links peninsula diplomacy to a broader effort to repair and normalize the relationship.
The source says Wang did not answer a reporter's question on a specific issue, leaving the limits of China's willingness to pressure Pyongyang unclear.

Open at SCMP Ask Claude ChatGPT

Xiaomi's quarterly revenue tops $15 billion as it funds AI and cars

# Signal Leiphone 12h

Xiaomi reported quarterly revenue of 108.9 billion yuan (roughly $15.3 billion), higher smartphone prices, and 104,199 vehicle deliveries while increasing R&D.

Operating profit reached 10.87 billion yuan (roughly $1.53 billion), up 104.6% sequentially, while phone average selling price rose 25.9% to 1,351 yuan (roughly $190).
Smartphone premiumization is being used to offset memory and component-cost pressure rather than relying solely on volume growth.
Automotive and other innovation revenue reached 24.9 billion yuan (roughly $3.5 billion), with vehicle deliveries up 28.2% year over year.
Xiaomi spent 9.2 billion yuan (roughly $1.3 billion) on research and development in the quarter, directing capital toward AI, operating systems, chips, robotics, and cars.
The company says its connected-device platform reached 1.16 billion devices, giving it a distribution base for AI services beyond phones.
The strategic question is execution: Xiaomi is attempting to turn a hardware ecosystem into a multi-cycle technology platform while entering capital-intensive automotive manufacturing.

Open at Leiphone Ask Claude ChatGPT

Chinese fighter designers warn that military AI can invent specifications

# Policy SCMP 11h

Engineers linked to China's J-20 and J-36 fighter programs warn that language models may hallucinate aircraft data and mislead defense analysis.

The warning comes from engineers at the Chengdu Aircraft Research and Design Institute, which the report identifies with China's advanced fighter programs.
The technical risk is not only incorrect text: a model could fabricate specifications, weapon performance, or causal explanations that look plausible to analysts.
The paper reportedly recommends treating language models as rapid information-mining tools while keeping human verification in the loop.
This is a notable Chinese defense signal because it frames AI reliability as an engineering and command problem, not merely a consumer chatbot issue.
Military adoption will depend on provenance, reproducible retrieval, domain-specific evaluation, and clear boundaries on when generated analysis may influence decisions.

Open at SCMP Ask Claude ChatGPT

Tencent's recommendation challenge tests whether one block can replace many

# AI QbitAI 18h

A Tencent-backed KDD Cup challenge drew 13,913 participants and tested whether a unified recommendation block can combine sequence modeling and feature interaction under millisecond latency constraints.

The industrial problem is architectural: traditional recommenders separate behavior sequences, user profiles, item attributes, and feature crosses, making GPU utilization and maintenance increasingly difficult.
The winning academic design used an intent token as a persistent interest state, while the industrial winner reportedly improved AUC by 0.0048 and cut per-sample compute by 18%.
Those small-looking efficiency gains matter at billion-scale ad traffic, where an extra fraction of inference cost is multiplied across every page view.
The contest's two tracks are valuable because offline accuracy alone is insufficient for a recommendation system that must respond within milliseconds.
The result suggests that recommendation systems may borrow the unified, scalable architecture idea from language models without inheriting their much looser latency budget.

Open at QbitAI Ask Claude ChatGPT

DeepSeek Harness grows from rough toolkit into agent marketplace

# AI GeekPark 23h

DeepSeek Harness reportedly reached 149,000 GitHub stars and 5,100 plugins in five days, as its modular architecture let the community repair its rough edges.

The project's central design makes nearly every subsystem replaceable, including model adapters, tools, file systems, sandboxes, scheduling, and even the agent loop.
That creates a trade-off: the initial product can feel unfinished, but external developers can add a desktop shell, image support, IDE features, and migrated workflows without waiting for the core team.
The architecture treats logs as the durable center, making model-visible context reconstructable and potentially improving auditability across interchangeable plugins.
The rapid plugin growth is a community metric, not evidence that the extensions are secure, reliable, or suitable for enterprise use.
The strategic lesson is that open agent systems can outsource product completion to developers, but they also shift permission, supply-chain, and maintenance risks into the ecosystem.

Open at GeekPark Ask Claude ChatGPT

MiniMax loses a senior engineering leader during its agent push

# AI QbitAI 17h

MiniMax's agent engineering head, Miao Yuhang, has reportedly left as the company expands coding, long-context, and multi-agent products.

Miao's reported remit crossed model engineering, agent infrastructure, coding, audio, and consumer multimodal products, making the departure more significant than a routine personnel change.
His public discussions emphasized harnesses, long-running tasks, and the engineering layer that turns a model into a reliable agent.
The timing comes after MiniMax launched agent teams, the M3 model, a coding product, and a video model, suggesting the company is entering a more execution-heavy phase.
The next destination is unknown, and the source relies partly on an employment-status update rather than a formal company announcement.

Open at QbitAI Ask Claude ChatGPT

China needs a stronger security offer for middle powers

# Policy SCMP 3h

An analysis argues that China must make its security partnerships more persuasive as US allies press regional states to align more closely with Washington.

The argument follows a sharp US push for partners to increase defense spending and accept a more demanding version of alliance cooperation.
China's challenge is presented as more than diplomatic messaging: middle powers need credible security benefits, not only criticism of US pressure.
The analysis suggests Beijing's economic relationships alone may not overcome concerns about coercion, regional risk, or dependence.
This is an interpretation rather than a new policy announcement, but it captures the strategic contest in which Southeast Asian states are trying to preserve room for maneuver.
The outcome will depend on whether China offers public goods and crisis reassurance that partners can explain domestically, not merely closer elite ties.

Open at SCMP Ask Claude ChatGPT

China's commercial-space sector moves toward repeatable operations

# Signal 36Kr 17h

A new industry report says China's commercial-space market reached 2.83 trillion yuan (roughly $399 billion) in 2025 and could exceed 3.5 trillion yuan (about $493 billion) in 2026.

The report identifies reusable launchers, flexible satellite production, and expanding remote-sensing and communications services as the transition from technology validation to scaled operations.
China recorded 50 commercial launches in 2025, while private and state-backed players are building manufacturing lines capable of producing satellites in the hundreds.
A reported 150 billion yuan (roughly $21 billion) of first-half 2026 financing and a rise in state capital participation to 59.8% show how heavily the sector depends on public-backed capital.
The most consequential technical milestone is reuse: the report cites both sea-based net recovery and a land recovery test for a launcher's first stage.
The market-size figures come from an industry report and are unusually large, so they should be treated as estimates whose definitions may include a broad supply chain.
The strategic direction is toward a domestic ecosystem spanning components, launch, constellation deployment, and AI-assisted data services.

Open at 36Kr Ask Claude ChatGPT

Chinese AI clouds treat reinforcement learning as a production system

# AI Leiphone 17h

A case study around GLM models argues that post-training reinforcement learning requires tightly coupled generation, environment, training, and state-management infrastructure.

The core loop is generate, execute against an environment, collect rewards, update the model, and generate again; unlike pretraining, the training data is produced live by inference.
The engineering bottleneck is matching generator throughput with trainer throughput while controlling rollout staleness, model-weight movement, KV-cache use, and environment state.
This reframes AI infrastructure from a collection of GPUs into a continuously operating production line for useful tokens and model capability.
The article cites GLM and DeepSeek deployments on Jiuzhang Cloud, but the commercial and performance claims come from the infrastructure provider's perspective.
For enterprises, the implication is that advanced agents may require specialized post-training and evaluation systems, not merely access to a larger base model.

Open at Leiphone Ask Claude ChatGPT

Chinese study warns against defense contractors dominating policy

# Policy SCMP 46m

A Shanghai Jiao Tong University study warns Beijing that defense contractors can distort strategy, using the US military-industrial complex as a cautionary example.

The study describes an American 'iron triangle' linking the Defense Department, private contractors, and congressional committees, and argues that profit incentives can expand military commitments.
Its audience matters: this is not simply an academic critique of the United States but a warning to Chinese policymakers about institutional capture and procurement incentives.
The framing reflects a recurring Chinese governance concern that private capital should serve state strategy rather than set it.
The report does not show that the study changes policy, but its publication provides a window into how Chinese strategists interpret US defense politics.
For technology executives, the relevant lesson is that Beijing is attentive not only to military capability but also to the organizational structure through which advanced technology is funded and deployed.

Open at SCMP Ask Claude ChatGPT

China briefing: youth unemployment, H200 approvals, and Wang Yi's diplomacy

# Analysis Sinocism 2h

A China-focused briefing links Wang Yi's Korea trip, 17.9% youth unemployment, Nvidia H200 purchases, and the JD.com regulatory clash.

The roundup is useful because it places diplomacy, industrial policy, technology controls, and labor-market stress in the same operating environment.
Its 17.9% youth-unemployment figure is a hard social signal, but the item does not provide the underlying statistical methodology or a time series.
The H200 reference suggests selective access to advanced foreign chips remains a policy variable even as domestic substitution is emphasized.
The briefing is a secondary digest, so individual claims should be checked against the cited government, company, or media sources before being used as evidence.

Open at Sinocism Ask Claude ChatGPT

Chinese AI firms are embedding models into complex systems

# AI Leiphone 21h

A pre-IJCAI analysis argues that China's AI competition is shifting from base-model scores toward agents, data loops, and integration with established industrial systems.

The article contrasts Alibaba's effort to give agents reusable skills with ByteDance's work on generating GUI-agent experience and training trajectories.
Baidu, Tencent, and Huawei are framed as putting models into search, maps, recommendation, knowledge graphs, and infrastructure rather than treating the model as a standalone product.
The important architectural point is that agents need tools, permissions, memory, evaluation, and business data; the model is only the cognitive layer.
This helps explain why Chinese platforms with large operational systems may compete through integration and distribution even when model differences narrow.
The analysis is forward-looking and does not establish that any of these approaches has achieved durable enterprise-scale economics.

Open at Leiphone Ask Claude ChatGPT

A Chinese embodied-data platform launches at roughly $720 per kit

# AI QbitAI 18h

51World launched a wearable data-capture and management system at 5,100 yuan (roughly $720), aiming to make robot training data cheaper and more standardized.

AperData combines first-person cameras, inertial sensors, synchronized capture, on-device quality checks, processing, evaluation, and dataset delivery.
The approach lets people perform tasks directly instead of occupying expensive robot hardware for teleoperation, potentially increasing collection parallelism.
The company claims physical trajectory consistency of 99% and more than tenfold data-production efficiency at equal cost, but these figures are not independently validated.
AperOne adds a reconstruction, training, evaluation, deployment, and operations layer for sites such as factories, mines, campuses, and power facilities.
The strategic issue is data quality, not raw video volume: detecting failed or unsynchronized segments at the edge can reduce bandwidth, storage, and cloud-compute waste.

Open at QbitAI Ask Claude ChatGPT

China leaves benchmark lending rates unchanged for a 15th month

# Policy Jiemian 15m

China kept its one-year and five-year loan prime rates at 3.00% and 3.50%, reflecting cautious monetary support despite softer investment and consumption.

The one-year rate remains 3.00% and the five-year rate 3.50%; the central bank's seven-day reverse-repurchase rate has also stayed at 1.40% since May 2025.
The unchanged fixing was expected because the policy-rate base and banks' wholesale funding costs had not materially moved.
The decision shows Beijing is using targeted and structural tools rather than broad rate cuts while headline growth remains within the official annual target range.
Analysts cited in the report expect further structural policy support and potentially new incremental measures near the end of the third quarter.
For technology companies, stable benchmark rates mean no broad financing reset yet; credit allocation, demand conditions, and local fiscal support may matter more than the headline LPR.

Open at Jiemian Ask Claude ChatGPT

Unitree's long road from garage robotics to public markets

# Analysis GeekPark 13h

A retrospective on Unitree argues that its success came from patient electric-drive engineering before humanoid robotics became a capital-market obsession.

The account traces the company to a small Hangzhou workshop and an early focus on making electrically driven quadrupeds more practical and affordable.
Its lesson for hardware founders is to solve technical and engineering constraints before building a grand commercial narrative.
The retrospective contrasts Unitree's early lack of investor interest with today's robotics boom, showing how timing and market fashion can obscure a viable technical route.
It is an investor's interpretation, not an independent operating history, but it offers useful context for why the company retained focus while the sector inflated.

Open at GeekPark Ask Claude ChatGPT

Nokia plans to sharply reduce its China operations

# Culture Solidot 18h

Nokia reportedly plans to cut most China-based staff and close factories by year-end, retaining mainly after-sales service.

The report says Nokia had about 7,200 employees across mainland China and the Hong Kong and Taiwan markets at the end of 2025.
A retreat from research, manufacturing, delivery, and supply operations would be more consequential than a simple sales-office closure because Nokia once maintained a broad local footprint.
The move illustrates how foreign technology firms are reassessing China operations amid tougher competition, supply-chain restructuring, and geopolitical uncertainty.
The source does not provide the affected headcount or a formal company announcement, so the scale and timing remain provisional.

Open at Solidot Ask Claude ChatGPT

At ChinaJoy, AI shifts from demo to game-production infrastructure

# Culture 36Kr 46m

ChinaJoy coverage shows Chinese game companies moving AI from marketing novelty into tools for development, rendering, distribution, and player-created content.

Tencent presented separate consumer game experiences and business tools, including natural-language game creation, asset production, rendering, and compression pipelines.
The industry discussion has shifted from whether AI can write a game to how teams should integrate it across planning, art, development, testing, and release.
This is a distinctly Chinese industry pattern: large platforms can connect game IP, creator tools, cloud infrastructure, and distribution in one commercial stack.
The report also emphasizes that AI is not replacing cultural production wholesale; Sony's showcase of Chinese and foreign games provides a counterexample to a purely AI-centered event.
The next constraints identified by practitioners are distribution, compliance, and content classification, not simply model capability.

Open at 36Kr Ask Claude ChatGPT

Experts say Golden Dome could intensify the China-Russia arms race

# Policy SCMP 4h

Experts warn that a US national missile-defense plan may offer incomplete protection while encouraging China, Russia, and North Korea to expand countermeasures.

The concern is a classic offense-defense dynamic: more missile defense can motivate adversaries to add penetration aids, volume, or new delivery systems.
China is a substantive part of the story because the system would alter strategic calculations across the same nuclear and missile environment in which Beijing is modernizing its forces.
The analysis does not establish China's planned response, so the immediate signal is incentive structure rather than a confirmed arms program.
For technology and aerospace companies, the relevant implications include demand for sensors, tracking, interception, and countermeasure systems, alongside tighter security scrutiny.

Open at SCMP Ask Claude ChatGPT

Seres reports $8.1 billion first-half revenue as Aito deliveries rise

# Signal Leiphone 12h

Seres reported first-half revenue of 57.493 billion yuan (roughly $8.1 billion), 70.07 billion yuan in R&D spending, and 10.2% year-on-year growth in Aito deliveries.

The company reported cash reserves above 73.15 billion yuan (roughly $10.3 billion), emphasizing balance-sheet resilience as Chinese electric-vehicle competition intensifies.
Aito's M9 surpassed 300,000 cumulative deliveries and reportedly led the 500,000-yuan vehicle segment for two consecutive months.
R&D spending rose 34.8%, with investment directed toward platform architecture, range-extender systems, safety, AI, digital applications, and robotics.
The figures are company-reported and the item does not provide profitability, margins, or delivery totals for the full Aito range.

Open at Leiphone Ask Claude ChatGPT

Taiwan's first locally developed satellite launcher fails its test

# Policy SCMP 13h

A Taiwan-developed launcher veered off course during its first test, alarming nearby residents who mistook it for an incoming missile.

The launch took place at a military test base operated by Taiwan's government-backed defense research institute, making the event relevant to the island's growing aerospace and missile capabilities.
The failure is an engineering setback rather than evidence of a weapon launch, but the confusion highlights how dual-use space technology is perceived in a tense security environment.
The report gives no technical cause for the deviation or recovery plan.

Open at SCMP Ask Claude ChatGPT

Chinese media expose fake employee discounts on flight apps

# Culture IT Home 30m

Chinese broadcasters found that supposed airline employee codes on travel apps usually unlock ordinary coupons, not secret fares, turning influencer content into a lead-generation funnel.

Reporters found that the codes led to themed pages promising bug fares and employee benefits, while normal searches returned the same ticket prices.
The coupons were real but generally worth only 20 to 40 yuan (roughly $3 to $6) and came with route, date, or usage restrictions.
An industry source described the tactic as channel marketing: platforms use the appearance of privileged access to drive users into app campaigns.
The case shows how Chinese online commerce converts trust in uniforms, insider language, and scarcity into measurable traffic, even when the underlying price advantage is minimal.
The report does not identify a regulator's enforcement action, so the immediate issue is consumer transparency rather than a new formal rule.

Open at IT Home Ask Claude ChatGPT

Unitree investor warns most embodied-AI startups will fail

# Analysis 36Kr 22h

An early Unitree investor argues that only a few embodied-AI companies will thrive because hardware engineering, supply chains, and commercial loops cannot be accelerated by funding alone.

The investor's forecast is opinion, not a market fact: he imagines roughly three companies doing very well, several surviving adequately, and most failing.
His framework evaluates robotics in three steps: can the technology work, can it be engineered into a reliable product, and can the business close a commercial loop?
He credits Unitree's electric-drive route, cost discipline, and control of core hardware for creating options across quadrupeds and humanoids.
The warning is useful against current valuation momentum, especially when prototypes are being priced as though manufacturing scale and demand were already proven.
The account also argues that Unitree's next moat may be an ecosystem of suppliers and capital partners rather than the robot body alone.

Open at 36Kr Ask Claude ChatGPT

ChinAI	403	0/0		HTTPError: 403 Client Error: Forbidden for url: https://chinai.substack.com/feed
ChinaTalk	200	1/20
Pekingnology	200	1/20
Ginger River Review	200	0/20
Sinocism	200	1/20
QbitAI	200	10/10
36Kr	200	8/30
Solidot	200	11/20
Leiphone	200	17/20
GeekPark	200	4/30
IT Home	200	12/60	+41 over cap
SCMP	200	10/50	+9 over cap
Initium	200	2/15
Jiemian	200	8/30
China Digital Times	200	1/15
The Wire China	200	0/50
Techmeme	200	0/15
Hacker News	200	0/20
V2EX	200	8/50	+24 over cap

Unitree took $48 billion out of a falling market, and its earliest backer says most of this sector dies

Unitree up 460%, Baidu down 12%

LandSpace landed a steel booster on legs

Nobody knows who runs the AI talks

DeepSeek's harness now ships whole agents

A Qianmen travelogue got deleted anyway

Threads we are pulling