China Desk — 19 Aug 2026, 20:38 UTC

A deleted post describes extreme security at Beijing's Qianmen gate

# China Digital Times 15h

A deleted Beijing travel account was censored after neutrally describing identity scans, pat-downs and bag checks near Tiananmen, illustrating how automated enforcement catches mundane speech.

The archived post reportedly described routine security procedures at Qianmen without criticizing the government or mentioning sensitive historical events.
The platform notice cited cybersecurity-law violations, suggesting either broad rule application or automated keyword and location filtering.
The significance is collateral censorship: a practical travel account can disappear even when it contains no explicit political argument.
The described security regime itself shows how politically sensitive urban space is managed through layered identity checks, physical searches and controlled movement.
Chinese online governance tends to define risk by context and location as well as by words, making apparently ordinary descriptions vulnerable near symbolic sites.
The episode also reveals a feedback loop in which users learn to avoid or euphemize sensitive places, reducing the public record of everyday state control.
Its importance lies less in one deleted post than in how low-level censorship normalizes self-censorship around national political landmarks.

Open at China Digital Times Ask Claude ChatGPT

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

# Pekingnology 1h

A newly translated 1998 speech shows Zhu Rongji urging CCTV's flagship exposé program to confront official abuse while retaining the Party's political line.

Zhu's remarks, delivered while he was premier, treated investigative television as an instrument for exposing local corruption and bureaucratic abuse, not as independent journalism in the liberal sense.
He described himself as an enthusiastic viewer and promoter of the program, while warning that officials exposed on air would resist it.
The speech reveals a recurring Chinese governance pattern: limited media criticism is encouraged when it disciplines lower-level administration and reinforces central authority.
The program's targets were local officials and agencies, leaving the central political system as the ultimate arbiter of what could be exposed.
Its publication now, alongside official mourning for Zhu, gives the text retrospective value as a reference point for comparing reform-era governance with today's tighter media environment.
The important distinction is between supervised accountability and adversarial scrutiny; Zhu was endorsing the former.
The wider significance depends on whether the speech is being presented as nostalgia for a more interventionist premier or as a usable model for current state-media management.

Open at Pekingnology Ask Claude ChatGPT

LandSpace recovers the first stage of its Zhuque-3 rocket

# SCMP 11h

LandSpace successfully placed a satellite in orbit and soft-landed Zhuque-3's first stage, giving China a flight-validated, land-recoverable orbital booster after the Falcon 9 model.

Also: Solidot

The 66-meter rocket launched from Jiuquan and landed about eight minutes later in Gansu after delivering the Honghu-03 satellite.
The report identifies nine Tianque-12 engines on the first stage and one Tianque-15 engine on the second stage, with a landing approach broadly similar to Falcon 9.
It is described as China's first land recovery of an orbital-class first stage using landing legs; an earlier Long March 10B demonstration used a sea-based net recovery.
Recovery is only the beginning of reusability: the hard economic test is inspection, refurbishment, turnaround time and repeat flight with a useful payload.
A successful booster landing lowers launch costs only when cadence rises and hardware survives multiple flights without expensive bespoke work.
The result sharpens competition between China's state launch programs and private companies such as LandSpace over reusable propulsion, launch infrastructure and satellite customers.
The next decisive milestone is a reflown stage and a sustained launch schedule, not the spectacle of a single landing.

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

Open at SCMP Ask Claude ChatGPT

Beijing blocks Chinese firms from helping an EU probe of JD.com

# SCMP 6h

Chinese coverage centers on Beijing's Justice Ministry banning assistance to an EU foreign-subsidy investigation, while Western framing emphasizes a failed JD.com takeover and extraterritorial conflict.

The ministry's order reportedly bars Chinese organizations and individuals from executing or assisting the EU investigation into JD.com's proposed acquisition.
Beijing characterizes the EU action as an improper use of extraterritorial jurisdiction, turning a competition case into a sovereignty dispute.
The move is more than rhetorical: it attempts to control access to corporate records, witnesses and evidence that a foreign regulator may need.
The clash illustrates how Chinese law can become a shield for domestic firms when overseas regulators investigate subsidies, state support or strategic acquisitions.
For multinational companies, the practical problem is dual compliance: cooperating with an EU inquiry may conflict with a Chinese prohibition on assistance.
Chinese coverage stresses legal sovereignty and protection of domestic enterprise, while Western discussion frames the episode as obstruction of an acquisition review.
The outcome will show whether Brussels can complete the case from public and non-Chinese evidence, and whether Beijing expands this tool to other foreign-subsidy investigations.

Open at SCMP Ask Claude ChatGPT

China and the US still lack a concrete agenda for planned AI talks

# SCMP 10h

With Xi Jinping's planned US visit approaching, the venue, lead agency, participants and scope of bilateral AI talks remain unsettled, showing how little institutional machinery exists for the dialogue.

The supplied report says Washington has not decided which part of the government should lead and that the two sides have not agreed on technical or private-sector participation.
That uncertainty matters because AI diplomacy spans export controls, safety evaluation, compute, model access and commercial competition, each owned by different US institutions.
A leader-level announcement without an agreed working channel risks producing symbolism rather than durable risk-management mechanisms.
China's likely preference is a state-to-state framework that treats AI as a strategic technology and sovereign capability; US agencies and companies may seek a more technical or operational format.
The absence of basic arrangements weeks before a planned visit suggests that strategic mistrust is limiting even low-level agenda-setting.
A useful dialogue would need narrow deliverables, such as incident communication, evaluation terminology or research-contact rules, rather than an attempt to resolve the entire technology relationship.
Whether the talks matter will be visible in who attends and whether the meeting creates a continuing channel below the leaders.

Open at SCMP Ask Claude ChatGPT

China plans a marine monitoring station near Scarborough Shoal

# SCMP 9h

Chinese authorities approved a marine observation station covering waters around disputed Scarborough Shoal, presenting a scientific facility that also strengthens Beijing's administrative presence.

The station was proposed by the Chinese Academy of Sciences' South China Sea Institute of Oceanology and approved on July 31, according to the supplied report.
Its stated mission is environmental observation and research, but the location gives China a durable platform for collecting data and maintaining institutional presence in contested waters.
Marine monitoring infrastructure can support fisheries, environmental enforcement, navigation and legal claims, so its strategic value exceeds the scientific label.
The move fits Beijing's pattern of using civilian and bureaucratic instruments to normalize control without announcing a new military base.
For the Philippines, the issue is not only the facility itself but whether Chinese research and monitoring activity changes the practical operating environment around the shoal.
Chinese coverage is likely to stress ecological stewardship and administrative legitimacy; regional critics will read the same infrastructure as incremental consolidation.
The effect will depend on construction, staffing, data publication and whether Chinese agencies use the station to support patrols or regulatory actions.

Open at SCMP Ask Claude ChatGPT

China prepares Chang'e 7 mission to search for lunar water ice

# IT Home 5h

Chang'e 7 has moved to the launch area for a lunar south-pole mission built around landing, hopping and probing permanently shadowed craters for water ice.

The mission targets the region south of 85 degrees latitude in the Aitken Basin, where permanently shadowed areas may preserve ancient volatiles.
Its planned capabilities include precision landing, legged surface travel, hopping and in-situ work inside permanently shadowed craters.
The scientific payloads will study surface conditions, regolith water and volatiles, geology, shallow structure, moonquakes, magnetic fields and the space environment.
The mission also carries international payloads, showing that China's lunar program can combine national control with selective scientific cooperation.
The strategic value is not just finding water; a successful south-pole operation would validate navigation, thermal management, communications and autonomous mobility needed for later infrastructure.
The supplied report frames Chang'e 7 as preparation for Chang'e 8 and a basic international lunar research station, making it part of a staged program rather than a standalone science flight.
The decisive evidence will be actual resource measurements and repeatable surface operations, not the launch narrative.

Open at IT Home Ask Claude ChatGPT

Chinese team uses a dark cavity to raise a superconductor's critical temperature

# IT Home 5h

A Chinese-led team reports the first experimental enhancement of superconductivity through vacuum fluctuations, raising a six-layer device's critical temperature by up to 5.4%.

The experiment placed niobium diselenide inside a terahertz split-ring resonator that reshapes the local electromagnetic environment without external photons.
The reported maximum increase in critical temperature was 5.4%, accompanied by higher critical current and critical magnetic field near the transition.
A resonance-shaped dependence on cavity frequency, plus control experiments, is presented as evidence that the effect is coupling rather than strain or sample degradation.
The proposed mechanism is virtual-photon exchange between the superconducting state and the cavity mode, which lowers the state's energy when the frequencies match.
This is a laboratory demonstration of a non-contact control knob, not a practical high-temperature superconductor or a device-ready architecture.
The result matters because it turns engineered electromagnetic boundary conditions into a possible materials-design parameter.
Its significance will depend on replication, absolute temperatures, effect size in other materials and whether the cavity can be integrated without sacrificing useful current or magnetic-field performance.

Open at IT Home Ask Claude ChatGPT

Chinese lab deploys humanoid robots for autonomous experiments

# AI QbitAI 14h

A Chinese national research laboratory is using Monte2 robots for long sequences of liquid handling, cell culture and toxicity testing, moving embodied AI from demos into scientific operations.

The robots reportedly perform reagent transfer, preparation, pipetting and instrument coordination without continuous human operation.
The system is built around an object-centered physical model that tracks samples, reagents, consumables, instruments and their relationships rather than executing one fixed motion.
The reported capability includes hours of continuous operation and sub-millimeter precision while adapting to liquid levels, tube angles and grasping forces.
The laboratory plans to expand the fleet to roughly 100 robots by the end of 2027 and coordinate them through a common AI system.
This is a valuable deployment domain because experiments generate measurable outcomes and repeated procedures, creating a natural feedback loop for models.
The important missing evidence is throughput, contamination rate, reproducibility against trained technicians and the cost of maintaining a humanoid platform.
If validated, the lab becomes both a customer and a data engine for physical AI, with implications for drug discovery and national research infrastructure.

Open at QbitAI Ask Claude ChatGPT

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

# AI GeekPark 8h

At Unitree's listing celebration, Wang Xingxing described a closed loop linking foundation models, code generation, simulation, real-robot testing and evaluation to accelerate physical AI.

The proposed loop is search for papers and code, generate training programs, run simulation, deploy to physical robots, score outcomes and retrain from the feedback.
Its key architectural claim is that AI should enter the entire development process, not merely serve as the robot's policy or language interface.
Simulation supplies cheap exploration, but real machines remain the scarce validation layer because balance, contact dynamics and task success cannot be fully trusted in a digital environment.
Unitree's reported installed base, including more than 30,000 quadrupeds and over 5,500 humanoids shipped in 2025, could provide a starting point for large-scale physical feedback.
The proposed moat is therefore not only a model; it is the combination of low-cost actuators, manufacturing, control software, deployment fleet and evaluation data.
That data will not automatically flow back from customer-owned robots: consent, standardization, privacy and heterogeneous operating conditions are unresolved.
The idea matters if Unitree can turn repeated real-world failures into reusable skills faster and more cheaply than competitors with better models but fewer deployed bodies.

Open at GeekPark Ask Claude ChatGPT

Tencent's recommendation contest tests a unified architecture at web scale

# AI QbitAI 12h

A Tencent-backed KDD Cup challenge drew 13,913 participants and showed a unified recommendation block improving AUC by 0.0048 while cutting per-sample compute 18% in the industrial track.

The contest asked teams to combine sequence modeling and feature interaction in one architecture instead of maintaining separate model branches.
The winning academic approach used an intent token that accumulated user and ad information through the network, while the industrial approach used a candidate query token as the interaction hub.
The industrial result matters because recommendation systems operate at millisecond latency and massive request volume, where a small compute reduction can outweigh a modest offline score gain.
The event attracted teams from 52 countries and regions, suggesting that Chinese internet companies are increasingly using open competitions to expose real production problems to global researchers.
The result does not prove a universal recommendation scaling law; it shows that architecture unification can be useful under a particular latency and feature distribution.
The strategic lesson is that industrial AI competition is shifting toward systems that improve accuracy and hardware efficiency simultaneously.

Open at QbitAI Ask Claude ChatGPT

Chinese researchers entangle quantum memories across 420 kilometers

# Signal Solidot 10h

A Chinese team reports entanglement between two cold-atom quantum memories over 420 kilometers of fiber, extending a key building block for metropolitan quantum networks.

The result reportedly exceeds the 230-kilometer threshold for direct, no-repeater entanglement distribution and connects matter-based qubits over a long fiber span.
Quantum memories are important because a network needs to store and synchronize entangled states rather than merely transmit photons once.
The work builds on the same team's reported progression from 50 kilometers in 2020 to a three-node metropolitan network in 2024.
Potential applications include repeater-based communication, distributed quantum computing and distributed sensing, but none follows automatically from the distance result.
The engineering questions are fidelity, rate, memory lifetime, synchronization and compatibility with deployable telecom infrastructure.
China's sustained lead in long-distance quantum-network demonstrations reflects patient state-backed investment in specialized laboratories and network testbeds.
The milestone matters more if the team can move from record distance to repeatable multi-node operation with useful entanglement rates.

Open at Solidot Ask Claude ChatGPT

Hong Kong's heat deaths expose a gap in climate adaptation

# Culture Initium 21h

Hong Kong recorded a 36.9-degree day and a 30.2-degree night as workers and older residents faced suspected heat deaths, prompting questions about whether the city treats heat as a public-health emergency.

The supplied analysis says the city endured nine consecutive days above 33 degrees and that the 30.2-degree minimum was its hottest night in 143 years.
It reports several suspected heat-related deaths, including older hikers and an outdoor renovation worker, illustrating the danger of persistent nighttime heat rather than isolated daytime peaks.
Hong Kong's dense, humid urban form magnifies exposure: cooling access, outdoor labor, housing conditions and social isolation matter as much as the official temperature.
The city's disaster-management culture is historically stronger on typhoons and rainstorms, while heat produces less visible disruption and weaker political urgency.
The labor dimension is central because outdoor workers may have limited ability to stop work, find shade or absorb lost income during warnings.
Useful adaptation would require targeted outreach, workplace enforcement, cooling centers and neighborhood-level monitoring, not only weather alerts.
The story matters as a governance test for an aging, high-density city where climate risk is increasingly a chronic service and labor problem.

Open at Initium Ask Claude ChatGPT

A 5,100-yuan embodied-data kit targets robot training bottlenecks

# AI QbitAI 12h

Chinese company 51World launched a 5,100-yuan (roughly $718) capture kit that records human demonstrations and filters bad data before upload, attacking the cost of robot data collection.

The AperData system combines first-person cameras, inertial sensors, hardware timestamps and a data platform for capture, quality checks, processing and dataset delivery.
Its key design choice is to have people perform tasks directly rather than occupy scarce robot bodies through teleoperation.
The company claims on-device filtering and synchronization can raise data-production efficiency by more than ten times at equal cost and achieve 99% physical consistency in delivered trajectories.
A companion application platform links reconstruction, simulation, evaluation, deployment and operations across sites such as factories, mines and campuses.
These are vendor claims, and the important missing metrics are task success after training, transfer across robot bodies and the cost of correcting bad demonstrations.
If the workflow works, it could make embodied-data collection resemble a scalable field-operations business rather than bespoke robotics research.

Open at QbitAI Ask Claude ChatGPT

China's space-computing bottleneck is getting data into orbit

# Analysis GeekPark 16h

Chinese space-computing startups are proliferating, but an 800MB model can reportedly take nearly four weeks to upload to orbit, exposing the infrastructure gap behind the hype.

The article distinguishes processing satellite-generated data in orbit from sending terrestrial datasets upward for AI training; the latter is the harder and newer business proposition.
The reported bottleneck is a low-rate command-oriented uplink, combined with short ground-station windows as low-Earth-orbit satellites circle the planet.
Downlink capacity has improved faster because large ground telescopes can receive and correct optical signals, while high-rate uplink remains an engineering challenge.
That asymmetry means orbital computing may be technically feasible yet operationally starved of models, datasets and software updates.
The Chinese market reportedly has more than 30 companies pursuing some form of space computing, showing how quickly an imported narrative can become an industrial policy and venture theme.
The first viable product may be orbital storage or edge filtering rather than general-purpose space data centers.
The decisive test is a repeatable, high-rate uplink service that can update models frequently enough to justify scarce satellite life.

Open at GeekPark Ask Claude ChatGPT

DeepSeek Harness turns a rough agent core into a plugin ecosystem

# AI GeekPark 18h

DeepSeek Harness reportedly grew from 50,000 to 149,000 GitHub stars in five days as developers supplied desktop shells, skills and interfaces, demonstrating the leverage of an open agent architecture.

The project makes the model adapter, tools, filesystem, sandbox, terminal, scheduler and even the main loop replaceable through plugins rather than reserving a privileged core.
Community developers quickly added desktop packaging, image support, IDE interfaces and migrated skills, turning missing product polish into an external development surface.
The architecture's commitment to exhaustive logs is significant: if the model saw it, the system should be able to reconstruct it, improving auditability and debugging.
The reported ecosystem exceeded 5,100 plugins and 3,500 authors, though repository metrics are a snapshot and do not establish sustained usage or security quality.
The strategic implication is that agent products may compete on extensibility and community repair speed as much as on model capability.
The same openness creates supply-chain, permission and quality-control risks; a plugin marketplace needs isolation and provenance or the ecosystem becomes an attack surface.

Open at GeekPark Ask Claude ChatGPT

Chinese fighter designers warn that military AI can invent specifications

# Policy SCMP 6h

Engineers linked to China's J-20 and J-36 programs warn that language models may hallucinate aircraft specifications, exposing a safety problem in AI-assisted intelligence work.

The reported paper says large language models can rapidly mine intelligence and analyze weapons performance, but may fabricate aircraft specifications and other technical facts.
For military analysis, a hallucinated parameter is not merely a bad answer; it can contaminate downstream simulations, threat estimates and procurement decisions.
The warning comes from an institute associated with Chengdu aircraft design, giving it unusual proximity to the engineering users who would need to validate model outputs.
The issue points toward retrieval, provenance tracking, numerical consistency checks and human sign-off rather than unrestricted model autonomy.
Chinese defense writing often presents AI as a force multiplier while emphasizing controllability and organizational discipline; this paper fits that pattern.
The practical architecture is likely a bounded analyst copilot with source-linked outputs, not a model trusted to generate authoritative specifications.
Its importance will rise if the warning is followed by procurement standards, red-team evaluations or documented deployment limits across military research organizations.

Open at SCMP Ask Claude ChatGPT

Chinese coastguard journal proposes robotic South China Sea outposts

# Policy SCMP 18h

A Chinese defense proposal envisions autonomous systems and machine swarms defending disputed islands, revealing how Beijing is thinking about persistent, low-cost maritime control.

The proposal reportedly replaces or supplements soldiers with coordinated machines designed to counter low-cost drone swarms and other asymmetric threats.
A robotic outpost could combine sensors, communications, unmanned surface or aerial vehicles and remote command, extending presence while reducing personnel exposure.
The concept fits a broader Chinese military trend toward distributed systems and autonomous cooperation rather than a few expensive platforms.
Because the proposal appeared in a journal supervised by a state shipbuilding group, it signals an institutional research direction but not an approved deployment plan.
The South China Sea setting matters: autonomous infrastructure could make contested features more continuously observable and harder for rivals to challenge without escalation.
The unresolved questions are communications resilience, rules for machine-initiated force, maintenance in harsh maritime conditions and control during spoofing or jamming.
Even if never built as described, the proposal shows how China is translating commercial robotics and drone technology into gray-zone maritime strategy.

Open at SCMP Ask Claude ChatGPT

Xiaomi uses higher phone prices and heavy R&D to ride out the cycle

# Signal Leiphone 7h

Xiaomi's second-quarter revenue reached 108.9 billion yuan (roughly $15.3B), while phone prices, cars and R&D all rose, showing a deliberate move from hardware volume toward a broader technology platform.

The supplied financial report says operating profit reached 10.87 billion yuan (roughly $1.53B), up 104.6% sequentially, while phone average selling price rose 25.9% to 1,351 yuan (roughly $190).
Phones priced above 3,000 yuan (roughly $423) made up 32.1% of mainland sales, indicating that Xiaomi is using mix and brand positioning to absorb memory and component inflation.
Smart electric vehicle and innovation revenue reached 24.9 billion yuan (roughly $3.5B), with 104,199 deliveries in the quarter and cumulative SU7 deliveries above 500,000.
R&D spending was 9.2 billion yuan (roughly $1.3B) in the quarter, up 18.9%, with capital expenditure directed heavily toward cars and AI.
The strategic bet is portfolio cross-subsidy: phones and internet services provide scale while vehicles, chips, operating systems, robotics and AI consume investment.
Xiaomi's connected-device platform reportedly reached 1.16 billion devices, giving it a distribution base for software and agent features beyond any single product.
The risk is execution complexity: premium phones, cars and frontier AI require different supply chains and support models, so rising revenue alone will not prove the platform strategy works.

Open at Leiphone Ask Claude ChatGPT

China's AI race shifts from bigger models to complex systems

# Analysis Leiphone 16h

A Chinese analysis argues that Alibaba, ByteDance, Baidu, Tencent and Huawei are embedding models into skills, trajectories, maps and business infrastructure rather than competing only on parameter count.

The article contrasts Alibaba's skill-library and computer-use path with ByteDance's focus on generating agent trajectories and experience through exploration, reflection and recovery.
Baidu's map and search infrastructure, Tencent's recommendation and graph systems, and Huawei's platform stack are presented as assets that models can absorb rather than replace.
The underlying thesis is architectural: an agent is a model plus tools, data, permissions, memory, evaluation and a feedback loop.
This is a more useful lens on China's AI market than model leaderboard comparisons because it asks who controls the surrounding systems where actions actually occur.
The analysis is broad and partly promotional, so individual product claims need separate validation.
The strategic question is whether China's incumbents can turn their installed data and workflow bases into defensible agent environments before model capabilities commoditize.

Open at Leiphone Ask Claude ChatGPT

Unitree's decade-long path from garage robots to public markets

# Analysis GeekPark 8h

A Chinese investor's retrospective portrays Unitree's rise as a decade of engineering discipline before the humanoid-robot boom, not a sudden product-market windfall.

The account says Unitree began with a low-profile electric quadruped effort when investors were focused on software and mobile-internet business models.
Its early appeal was technical: a compact, product-like electric robot that contrasted with much larger and more expensive laboratory machines.
The narrative emphasizes founder Wang Xingxing's combination of intense technical curiosity and deliberately cautious commercial judgment.
The retrospective says the early market was primarily research institutions, a narrow but plausible beachhead while hardware and supply chains matured.
That history helps explain why Unitree's current valuation is tied to manufacturing and engineering credibility rather than only to a humanoid-robot narrative.
The source is an investor-affiliated account and therefore offers useful context but should not be read as an independent operating history.
The strategic lesson is that hardware companies can survive a cold financing cycle when they identify a technically reachable market before the larger story becomes fashionable.

Open at GeekPark Ask Claude ChatGPT

Nokia plans a major retreat from manufacturing in China

# Culture Solidot 13h

Nokia reportedly plans to cut most China-based staff and close factories by year-end, preserving mainly support operations and underscoring the restructuring of foreign telecom supply chains.

The supplied report cites about 7,200 employees across mainland China and the Hong Kong and Taiwan regions at the end of 2025, though the affected number is not disclosed.
Former employees reportedly describe compensation based on service years plus three months of salary, but the strategic fact is the withdrawal of manufacturing and much of the local footprint.
Nokia's China organization had included research, marketing, services, global delivery and supply, so the move may affect more than factory labor.
The retrenchment reflects a market in which domestic vendors dominate much of China's communications equipment ecosystem and foreign vendors face geopolitical and commercial pressure.
The remaining service presence would preserve customer support while reducing fixed assets and exposure to a difficult local market.
This is a labor and industrial-policy signal as much as a corporate restructuring story: localization can deepen even as foreign suppliers remain present in narrow technical roles.

Open at Solidot Ask Claude ChatGPT

Chinese video creators are replacing prompt roulette with 3D previs

# AI QbitAI 13h

A Chinese creator workflow uses AI-generated white-box 3D scenes to control camera paths and blocking before rendering video, addressing a real weakness in text-only direction.

The proposed workflow builds a rough scene, places cameras and characters, records a previs pass, then feeds that spatial reference to video models.
The technical advantage is explicit geometry: depth, occlusion, relative distance and motion paths are represented directly instead of being translated twice through prose.
The approach resembles earlier control methods that supplied depth or pose maps while leaving texture and lighting to the generator.
It does not eliminate expertise; creators still need to decompose shots, define movement and understand scene hierarchy.

Open at QbitAI Ask Claude ChatGPT

Extreme quantization pushes large language models toward one-bit weights

# AI Leiphone 10h

A researcher moving from ETH Zurich to Hong Kong Polytechnic University describes post-training methods that preserve usable behavior at one and two bits, targeting the memory wall in edge AI.

The reported BiLLM method isolates sensitive weight channels and quantizes most remaining weights to one bit without full retraining, while SqueezeLLM uses groupwise allocation at two bits.
The systems problem is memory bandwidth as much as arithmetic: lower precision reduces model storage and data movement, which often dominate inference on small devices.
The work identifies a long-tail structure in weights where a small number of channels carry disproportionate sensitivity, allowing precision to be spent selectively.
The source claims a 7B FP16 model would occupy about 14GB, illustrating why aggressive compression matters for phones, vehicles and medical devices.
Below two bits, the remaining obstacles include logical-reasoning degradation, weak native hardware paths, activation quantization and KV-cache size.
The research is not China-originated, but the Hong Kong appointment gives it an institutional China dimension and creates a potential bridge between European research and Asian deployment.
Its practical importance depends on end-to-end latency, kernel support, accuracy across long-context and tool-use tasks, and whether hardware vendors expose efficient one-bit paths.

Open at Leiphone Ask Claude ChatGPT

China's commercial space sector shifts from tests to operating scale

# Signal 36Kr 12h

A Chinese industry report says commercial space revenue could exceed 3.5 trillion yuan (roughly $493B) in 2026, with reusable rockets, satellite factories and state capital driving expansion.

The report estimates 2025 market size at 2.83 trillion yuan (roughly $399B), up 21.7%, and cites 50 commercial launches in that year.
It describes a shift from bespoke satellites toward standardized, flexible production, with some leading manufacturers targeting annual output in the hundreds.
Reusable launch is moving from single recovery demonstrations toward reflights, while launch sites and constellation deployment become the system bottlenecks.
The report says first-half 2026 financing exceeded 15 billion yuan (roughly $2.1B) and state-linked capital's share rose to 59.8%, revealing how public money is shaping the private sector.
The headline market-size estimate is unusually large and source-dependent; it should not be confused with revenue captured by launch firms or satellite operators.
The strategic picture is an industrial policy model: state programs set infrastructure and demand conditions while private companies specialize in components, launch and applications.
The next proof point is recurring revenue from satellite services, not more company formation or one-off financing.

Open at 36Kr Ask Claude ChatGPT

Wang Yi and South Korea's foreign minister seek a reset

# Policy Jiemian 5h

Chinese coverage presents Wang Yi's Seoul talks as a bid to stabilize ties, deepen trade and technology cooperation, and keep Korea aligned but not strategically captive.

Wang called for frequent high-level contact, deeper economic integration and progress on the second phase of the bilateral free-trade agreement, according to the supplied readout.
The proposed growth areas include artificial intelligence, green transition, digital economy, advanced equipment and services for an aging population.
The Chinese message also urged Seoul to develop foreign relations independently, a familiar formulation that signals resistance to tighter alignment with Washington and Tokyo.
South Korea's foreign minister endorsed continued high-level exchanges, the one-China policy and cooperation in trade, high technology and artificial intelligence.
The timing before the countries' 35th anniversary and a planned regional summit gives both governments a calendar for symbolic normalization.
The central tension is economic interdependence versus security alignment: Beijing wants commercial and political recovery without demanding a formal rupture with Seoul's allies.
Whether the reset matters will depend on concrete trade negotiations and whether technology cooperation survives pressure from semiconductor and security restrictions.

Open at Jiemian Ask Claude ChatGPT

China's week in view: Zhu's funeral, Korea ties, housing and consumption

# Analysis Sinocism 21h

A Chinese-focused roundup pairs Zhu Rongji's funeral with Xi's diplomacy, Unitree's IPO and economic-policy signals, offering a useful map of what Chinese observers considered the day's important stories.

The roundup says top leaders attended Zhu's funeral while former leader Hu Jintao sent a wreath, framing the event as an orderly end-of-era ritual rather than a political contest.
It places South Korea outreach, housing provident-fund changes, consumption and lower-tier markets alongside technology and defense, revealing a domestic agenda concerned with both stability and demand.
As a newsletter synthesis, it is valuable for editorial framing but should not replace the individual reports for facts about any one event.

Open at Sinocism Ask Claude ChatGPT

Seres reports 57.493 billion yuan in first-half revenue

# Signal Leiphone 7h

Seres reported 57.493 billion yuan (roughly $8.1B) in first-half revenue and 70.07 billion yuan (roughly $9.9B) in R&D spending, with Aito deliveries up 10.2%.

The supplied report says Seres held more than 73.15 billion yuan (roughly $10.3B) in cash at the period end and that Aito's M9 surpassed 300,000 cumulative deliveries.
R&D spending rose 34.8%, reflecting continued investment in vehicle platforms, range-extender systems, intelligent safety and robotics.
The figures show a Chinese automaker using Huawei-linked premium positioning to fund a broader technology and product expansion.

Open at Leiphone Ask Claude ChatGPT

Unitree investor warns only a few humanoid firms will thrive

# Analysis 36Kr 17h

An early Unitree investor argues that perhaps three embodied-AI companies will thrive, ten will survive, and the rest will fail, a useful counterweight to China's robot-investment frenzy.

The investor's test for a robotics company is whether its technology works, engineering can scale and a commercial loop closes; funding and valuation do not substitute for any of the three.
He credits Unitree's early electric-drive choice, cost discipline and control of its own hardware for creating optionality across quadrupeds and humanoids.
The warning is about capital concentration: hardware markets with long iteration cycles are unlikely to support dozens of similarly positioned winners.
The interview is an investor's judgment, not a market forecast, but it captures a growing concern that abundant financing is pulling schedules and valuations ahead of deployment economics.

Open at 36Kr Ask Claude ChatGPT

A physics dialogue explores low-power materials that can remember

# Culture QbitAI 14h

A Chinese science program uses soft-matter physics to discuss disordered materials, jamming and possible low-power memory-like computation, bridging public science and AI speculation.

The discussion explains jamming as a transition in which densely packed particles acquire rigidity without forming a conventional crystal.
It connects soft matter and nonequilibrium systems to materials whose collective states might be tuned, stored and used for computation.
The episode is an interview and educational synthesis, not a report of a demonstrated intelligent material or a product milestone.

Open at QbitAI Ask Claude ChatGPT

Mongolia apologizes after dispute over a Chinese-backed oil site

# Policy SCMP 14h

Mongolia apologized after Beijing objected to protesters at a Chinese-invested oil exploration site, showing how economic projects can become bilateral sovereignty and security issues.

The supplied report says protesters entered the exploration area, confronted Chinese workers and streamed the incident online.
A senior Mongolian official apologized after the Chinese embassy raised concerns, indicating that Beijing treated the event as a diplomatic protection issue rather than only a local dispute.
The episode illustrates the political exposure of Chinese overseas investment: project security, local consent and nationalist backlash can quickly move into state-to-state channels.
The report does not establish whether the exploration project itself violated local rules or how broad public opposition is.

Open at SCMP Ask Claude ChatGPT

China's tech brief: Unitree lists as chip workers receive large equity awards

# Signal Leiphone 20h

A Chinese roundup mixes Unitree's listing, Cambricon employee equity, vocational training and consumer-business changes; its value is breadth, not a single decisive event.

Also: 36Kr

The roundup says 124 Cambricon employees will receive shares worth about 690 million yuan (roughly $97M) at the cited market price, illustrating how the domestic chip boom is reshaping labor incentives.
It also reports Unitree's 150.80-yuan (roughly $21) issue price and the presence of major technology and state-linked investors, but repeats coverage elsewhere in this batch.
The item is useful as a snapshot of Chinese business mood, where robotics, semiconductor wealth, industrial training and small consumer fees appear in the same news cycle.

China's morning tech roundup tracks agents, robotics and industrial policy 36Kr

Open at Leiphone Ask Claude ChatGPT

MiniMax loses a core engineering leader during its agent push

# Analysis QbitAI 12h

MiniMax's agent engineering head, known for work spanning models, harnesses and products, has reportedly left as the company deepens its coding and long-context strategy.

The supplied report identifies the departing executive as responsible across MiniMax's model, coding, agent, audio and consumer application lines, although the next destination is unknown.
His role illustrates how large-model companies increasingly depend on engineering leaders who connect base models to training systems, harnesses and products.
The departure is a personnel signal, not evidence of a strategic reversal; the report provides no reason for the exit or indication of impact on projects.

Open at QbitAI Ask Claude ChatGPT

Forum user alleges malicious telemetry in a Chinese open-source admin project

# Signal V2EX 7h

A Chinese developer alleges that packages used by gin-vue-admin inject hidden telemetry and licensing prompts through npm, a serious but unverified software-supply-chain warning.

The forum post alleges that two dependencies hide code that reads project variables, calls an external image endpoint and displays authorization prompts when network access fails.
It further alleges anti-analysis behavior that skips localhost and headless-browser environments, plus a second package that can blank the generated site if the first package is removed.
These are claims from one investigator's code review, not established facts; the post cites third-party scanner warnings but no maintainer response or independent reproduction.
If confirmed, the pattern would blur the line between license enforcement, telemetry and malicious build-time interference.
The case illustrates why dependency review must include transitive packages, generated artifacts and network behavior rather than only the top-level repository source.
The reported protocol and package names should be treated as indicators for immediate review, not as a definitive attribution of intent.

Open at V2EX 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	6/30
Solidot	200	11/20
Leiphone	200	17/20
GeekPark	200	4/30
IT Home	200	12/60	+45 over cap
SCMP	200	10/50	+6 over cap
Initium	200	3/15
Jiemian	200	8/30	+11 over cap
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