A Hong Kong University of Science and Technology and CMU-linked team, with Chinese company Rockchip Matrix, released EdgeRock-4B, claiming near-constant state memory for long reasoning and much higher single-GPU throughput.
More
- The report says the model holds its state at roughly 72 MiB across contexts up to 256K tokens, a 99.1% reduction from its reference model; a separate benchmark configuration reports 97.76 MiB per sequence.
- On one 80 GB GPU, the team reports 12,124 output tokens per second at 512 concurrent sequences, versus 1,889 for the Qwen reference model, a 6.42x difference.
- The architectural aim is not simply to compress each token’s KV cache but to avoid memory growing with sequence length while retaining long-range reasoning information.
- The article reports an average 82.9% pass rate on AIME 2024–2026, far above the listed bounded-state baselines; these results are from the releasing team and need independent replication.
- If the memory and quality tradeoff holds under independent evaluation and diverse workloads, fixed-state inference could substantially raise concurrency and make long-context workloads more predictable on constrained hardware.
- The provided material does not establish performance across hardware generations, production serving stacks or real agent workloads, and benchmark comparisons may depend on task and tuning choices.
Unitree's CEO predicts a near-term embodied-AI inflection and describes a robot self-improvement loop, while the company says it has produced roughly 10,000 to 20,000 G1 humanoids.
More
- Wang Xingxing said robots might soon handle about 80% of tasks across 80% of unfamiliar settings when given spoken or written instructions; this is his forecast, not an independently verified benchmark.
- He identified the alignment between model inputs and physical actions as a major unresolved challenge, even as he argued that planning is largely solved.
- Unitree's proposed self-evolution loop combines perception, decision, execution, evaluation and learning, with progress depending on stronger base models, diverse data, real-world deployments and reusable skills.
- Wang said the G1 has had roughly 10,000 to 20,000 units produced since its May 2024 release, a meaningful manufacturing signal but not a reported sales or active-deployment figure.
- The company also showed a new humanoid and a partner ecosystem spanning compute, components, research and applications.
- The strategic question is whether deployed robots can produce reliable, repeatable data and transferable skills, rather than merely expanding demo capability.
Hong Kong coverage surveys September rulings on national-security convictions, union rights and alleged sedition, showing how courts are drawing boundaries around political speech and civic organizing.
More
- Three leaders of the Alliance in Support of Patriotic Democratic Movements of China received prison terms of five years and two months to seven years and three months; the organization was fined HK$1.5 million.
- The court classified the sedition case as serious despite finding no violence, concrete implementation plan or sufficient evidence of actual harm, citing the defendants' influence and sustained activity.
- In a separate case, a court found the Wall Street Journal's parent company guilty of obstructing union rights but not unlawful dismissal, after journalist and union leader Ronson Chan brought a private prosecution.
- A former Stand News editor is appealing his sedition conviction; prosecutors argued that press freedom does not protect publication of material judged to have seditious intent.
- The article also reports that an appeal by trustees of the 612 Humanitarian Relief Fund was rejected, with the court treating the fund as a local society subject to registration rules.
- Together, the cases show legal exposure across political advocacy, journalism and labor organizing, not only direct protest activity.
The Party's four-day Central Committee meeting opened with official coverage stressing discipline, anti-corruption and loyalty to Xi, signaling that political control is central to the agenda.
More
- The fifth plenum of the 20th Central Committee began in Beijing, with a Politburo work report among the listed agenda items.
- State media linked the country's success to adherence to Party leadership under Xi Jinping, giving the meeting a strong loyalty-and-discipline frame.
- The emphasis comes as the Party uses internal discipline and anti-corruption campaigns as tools of governance and cadre control.
- The report describes the opening and official messaging, not specific policy decisions or the meeting's final conclusions.
- What matters next is whether the plenum produces concrete governance or economic priorities beyond reaffirming political discipline.
A report on local-government investment describes cities chasing embodied-AI projects to satisfy industrial targets, while founders divide operations across regions to match local funding and investment-attraction priorities.
More
- The report cites more than 300 embodied-AI financing events in China in the first half of 2026, totaling 93.5 billion yuan (roughly $13 billion); state-backed investors participated in 42% of large deals.
- Local governments see robotics as easier to localize than fields requiring rare national research infrastructure: components, robot bodies, software and data operations can be placed in different cities.
- Some startups reportedly split headquarters, research, production lines and data centers among jurisdictions to fit each locality's investment priorities.
- Interviewees describe projects being packaged around investment-attraction deliverables such as factory sites, equipment purchases and announced orders, sometimes ahead of evidence of durable commercial demand.
- The report says falling robot-company valuations and tighter listing expectations have weakened the easy exit narrative for later-stage investors.
- This is a portrait of how local state capital can shape industrial geography as much as fund companies; the test is whether supported clusters generate repeat customers and productive capacity.
Chinese coverage frames optical chips as an AI-infrastructure bottleneck, but warns that US restrictions and difficult materials processing could blunt the domestic-substitution opportunity.
More
- Optical interconnects help address the bandwidth and power limits of moving data among AI accelerators, making lasers and other optical chips increasingly valuable in high-speed modules.
- The article says optical chips can account for more than 30% of optical-module cost, with the share potentially rising as products move from 800G toward 1.6T and 3.2T.
- Unlike conventional silicon logic, high-speed optical devices depend heavily on materials, epitaxial growth and process control; the article compares the manufacturing challenge more to silicon carbide than to shrinking digital transistors.
- A full indium-phosphide production line can require investment of tens of billions of yuan (several billion dollars), while imported equipment, narrow process windows and low yields raise barriers to entry.
- The analysis says Chinese firms are stronger in module assembly than in upstream optical chips, leaving the component bottleneck exposed to US foreign-direct-product-rule risk.
- The opportunity depends on yield learning, customer qualification and export-control scope; the article’s market and margin projections are estimates, not demonstrated outcomes.
After a six-win World Superbike season, Zhang Xue used a 40.74-kilometer water-skimming stunt in his hometown to celebrate a Chinese motorcycle maker’s rapid rise from grassroots startup to export contender.
More
- The article links the company’s racing results to a push to overcome the long-standing perception that Chinese motorcycles are low-cost assemblies rather than high-performance machines.
- It says Zhang invested 9.33% of 2025 output value—about 69.6 million yuan (roughly $9.8 million)—in engine development and international racing.
- The race rules require competition bikes to be based on production models, giving results a potential credibility effect with buyers and dealers, though racing success does not by itself prove long-term reliability.
- The report says the company’s 820RR has 97% domestic content and that its engine, chassis and control systems were developed in-house; these claims are not independently audited here.
- The story illustrates a founder-led strategy that treats motorsport as a costly product-validation and brand-building channel in a sector where suppliers and customer trust have historically favored foreign incumbents.
- The reported financing and valuation surge after racing wins show how quickly attention can translate into capital, but do not establish durable margins or repeatable export demand.
Chinese coverage emphasizes Kimi K3.1 Pro's benchmark results and claimed architectural progress, while US coverage focuses on allegations about undisclosed paid influencer promotion.
More
- The Chinese article cites an Artificial Analysis intelligence ranking of fourth overall and a 65.7 score on Terminal-Bench 4, just below the score it attributes to Claude Opus 5.5.
- It argues that a flagship Pro release after a stretch of lighter models counters doubts about whether Chinese labs can sustain frontier progress, but its framing is openly promotional.
- Moonshot's previous K3 architecture is described as using a second-order optimizer, hybrid linear attention, attention residuals and sparse experts; the article presents these as original innovation, not independent verification.
- The piece claims the open-weight versus closed-model intelligence gap could shrink to three points if weights are released, while that condition means the comparison is not yet settled.
- Whether this marks repeatable research progress depends on reproducible results, released weights and performance beyond selected benchmarks.
A 100-minute AI-assisted film about Sanxingdui opened in Chinese cinemas with low initial box-office results, exposing the distance between cheaper generation and cinema-ready production.
More
- The film reportedly cost tens of millions of yuan (several million dollars) to make and grossed about 1.18 million yuan (roughly $170,000) by October 25.
- Its production used nearly 100 people and reportedly generated more than 1.2 million source images for about 2,000 shots; the team manually corrected frames to maintain character and scene consistency.
- That labor profile complicates the claim that generative video makes feature filmmaking cheap: it reduces some location and visual-effects costs while shifting effort into selection, continuity and correction.
- The project began as a lower-cost test after an AI short on the same subject reportedly drew a large online audience, illustrating how Chinese studios are using short-form platforms to validate concepts.
- The weak early ticket sales are a useful counterweight to industry excitement about AI film production; the source does not establish whether the release can recoup its costs or attract a durable audience.
A Chinese court found that a device advertised as refreshing and grabbing delivery orders in five milliseconds distorted competition among couriers and ordered its seller to pay 3 million yuan (about $420,000).
More
- The case puts platform fairness and worker competition at the center of an intellectual-property and unfair-competition ruling, rather than treating the tool as merely a nuisance to platform operators.
- Couriers said the device helped users capture better routes and higher-value jobs, leaving ordinary riders at a disadvantage.
- The court said the product also raised platform maintenance and enforcement costs and could harm merchants and consumers indirectly.
- The report notes similar order-grabbing tools in ride-hailing, freight and ticketing, suggesting a recurring conflict between platform allocation systems and workers seeking an edge.
- The ruling may strengthen platforms’ ability to pursue sellers of automation tools, but the report does not explain how broadly the legal theory will apply or whether affected couriers face penalties.
Primotion’s Paimon-0 ranked first on several robot benchmarks, while the company says its FPGA control path cut EtherCAT timing error from 30 microseconds to 152 nanoseconds.
More
- The reported control result compares software scheduling with FPGA hardware logic at a 1 kHz control cadence; the shift targets jitter by moving timing-critical work below the CPU scheduling layer.
- The company’s proposed stack spans model training and adaptation, deployment tools, cloud planning, local action control, runtime coordination and feedback-based evaluation.
- Its architecture divides work between a cloud planner for open-ended tasks and a local controller that uses visual, tactile and force feedback for fast closed-loop execution.
- The article says Paimon-0 led RoboColiseum, RoboCasa365 and VLA-Arena results, but those benchmark claims are reported by the company and lack independent validation here.
- Starting with models and real customer tasks, then deriving chip requirements, is a notable alternative to designing a robotics chip before the workload and compute split are understood.
- The company is only three months old, and leaderboard results do not yet establish reliable deployment across varied robots, factories or safety-critical tasks.
iFlytek’s Spark-X2.6 combines a large mixture-of-experts model with smaller local models, routing tasks between on-device and cloud execution based on capability, privacy and cost.
More
- The company says Spark-X2.6 has 293 billion total parameters, activates about 30 billion per inference and supports a one-million-token context; it is offered through iFlytek’s MaaS platform.
- Its proposed routing has three modes: local execution, cloud execution after on-device redaction, or a split workflow in which the cloud guides and the device acts.
- The release positions domestic compute as a full-stack effort spanning training, inference and edge deployment, where software adaptation and cluster stability matter alongside chip availability.
- iFlytek claims a fourfold efficiency improvement on domestic training hardware, 95% hit rates for a four-layer KV cache and lower token consumption in about 60% of high-frequency cases; the report gives no independent verification.
- The architectural tradeoff is meaningful for regulated or privacy-sensitive workloads: local processing can constrain data movement, while cloud access preserves access to larger models.
- The article is a product demonstration and does not establish whether the routing controls, claimed savings or model quality hold across customer environments.
Chinese coverage foregrounds Anthropic's revenue surge and headline loss, while US coverage focuses on its unusual governance; the filing says most of the reported first-half loss came from noncash financing revaluation.
Also:
TMTPost, Techmeme
More
- Anthropic's S-1 reports $16.3 billion in first-half 2026 revenue and an $80.7 billion net loss; the Chinese article says about $72.4 billion of that loss reflected fair-value changes in financing instruments.
- The company reported $894.5 million? No: $8.945 billion in compute and infrastructure costs in the first half, a large operating expense even after excluding financing revaluation.
- The filing also describes rapid enterprise adoption, including more than 1,900 customers with at least $1 million in trailing annual revenue and net revenue retention near 900%, as reported by the article.
- Chinese coverage emphasizes the scale of the financial figures and the potential IPO valuation, whereas the US Techmeme item highlights the governance design for a mission-driven public company.
- The headline loss should not be mistaken for equivalent cash burn: the distinction between noncash accounting charges, operating losses and future infrastructure commitments is central to evaluating the business.
A Chinese robotics partnership says its four-legged robots can map steelworks, navigate around temporary obstacles and collect equipment data, with deployments reported at several plants.
More
- The system combines a Unitree robot, DeepMirror spatial cameras and navigation software, a task platform and optional thermal, acoustic or gas sensors.
- The workflow links mapping and localization to route planning, obstacle avoidance, inspection and archived alerts, rather than treating locomotion alone as the industrial task.
- The report says deployments at Xiangtan Steel and Jinnan Steel have reached routine operation, but gives no uptime, detection accuracy or labor-savings data.
- This division of labor—general-purpose robot hardware plus an industry integrator responsible for sensing, navigation and workflow—is a plausible route to industrial deployment.
- The company says the approach can transfer to other hazardous inspection environments, but each site still requires adaptation to its layout, safety rules and maintenance process.
Chinese coverage reads this year's Programmer's Day as a labor-market identity crisis: AI has widened the celebration from coders to anyone who builds with AI, while entry-level work appears to be narrowing.
More
- The October 24 observance began as an informal in-group reference to 2 to the 10th power, then became a large public event increasingly centered on AI and the broader idea of building software.
- Online posts ranged from joking that AI deserved the holiday's golden keyboard to selling used programming books under a label meaning 'old-school coding.'
- The article cites domestic data showing backend developer job demand down 41% and AI jobs up 7.9 times, while noting that other evidence and industry voices differ on entry-level hiring.
- It describes a ladder being compressed at its lowest rung: AI may take on routine implementation while newcomers are expected to review generated code, understand legacy systems and coordinate work.
- That shift matters beyond headcount: junior engineers traditionally learn through implementation, so replacing those tasks could weaken the path by which future senior engineers acquire judgment.
- The story is a cultural reading of a changing profession, not proof that programming work is disappearing; the evidence on hiring and productivity remains mixed.
The partners released a conversational agent based on the Three-Body universe, using separate generation, knowledge-retrieval and consistency-checking steps to keep the character on-model.
More
- Three Body Universe defines the character and knowledge boundaries, while Wuwen provides agent development, memory management and inference infrastructure.
- The system separates character expression from retrieval so reference material does not simply overwrite the voice of the character; an independent review stage flags lore conflicts and out-of-character responses.
- The companies describe a generate-check-revise loop with traceable sources and human review for higher-level judgments about tone and dramatic quality.
- The infrastructure uses a split inference design with prefix caching; the company reports a 51.5% reduction in time to first token and 37.5% lower per-token cost.
- This is a practical example of turning fictional-character consistency into a system-design and evaluation problem, though the reported performance figures come from the provider.
Moonshot says K3.1 Pro is a post-trained version of its existing large MoE model, with stronger coding results and lower token prices; its promised open weights could make the release especially consequential.
More
- The model retains K3's reported 2.8-trillion total parameters, 104-billion active parameters and one-million-token context window; the change is in post-training, not a new base architecture.
- Moonshot says it rebuilt parts of its reinforcement-learning framework, sandbox infrastructure and long-context reasoning stack to train across hundreds of downstream tasks.
- On Terminal-Bench 4.0, the article reports 65.7%, close to Claude Opus 5.5's 66.4%; the previous K3 score cited from Artificial Analysis was 13%, though benchmark setups may differ.
- The reported API price is 13 yuan per million input tokens and 68 yuan per million output tokens, roughly $1.80 and $9.60 respectively, below K3 pricing; third-party estimates say the model may use more output tokens per task.
- The release also emphasizes 3D tool use and financial reasoning, while user feedback cited in the article is mixed on visual design quality.
- If the full weights are released as promised, the combination of a large-context MoE model and competitive coding performance would give developers more scope to inspect, adapt and self-host a leading Chinese model.
A Chinese report says ForgeKernel, whose operators were generated and iterated by AI, topped NVIDIA- and Ascend-focused benchmarks in the same week DeepSeek open-sourced its inference stack.
More
- DeepSeek released DeepInfer, extending its open-source inference work beyond model weights and operator libraries to a broader model-serving path.
- The article says a human-tuned FlashMLA operator had reached more than 90% of hardware performance on some chips, then contrasts it with ForgeKernel's reported benchmark results.
- ForgeKernel reportedly ranked first on both the SOL-ExecBench and FlashInfer-Bench set and Ascend CANNBench A3, with a 1.835x speedup and a 25% lead over the next entry on the latter.
- The report also says AI iteration surpassed DeepSeek's open NVIDIA operator on a DeepSeek V4.1 Flash workload; it provides no independent methodology or reproducibility details in the excerpt.
- The strategic point is that kernel optimization may become another model-assisted layer of the compute stack, potentially improving utilization across both NVIDIA and domestic accelerators.
Official yearbook figures show people aged 65 and over exceeded children in the dependency ratio in 2025, underscoring the scale of China's aging and shrinking-household challenge.
More
- The 2026 Statistical Yearbook puts the elderly dependency ratio at 23.1% in 2025, above the child dependency ratio of 22%.
- People aged 65 and over numbered 223.65 million, or 15.9% of the population; children aged 0 to 14 numbered 212.76 million, or 15.1%.
- The elderly ratio rose from 15% in 2016 to above 20% in 2021, showing a rapid shift in the age structure rather than a one-year anomaly.
- Average household size fell to 2.52 people, from 3.44 in the 2000 census, changing the family networks available for caregiving.
- The combined pressures affect pensions, health care, labor supply and consumer demand; the figures do not by themselves reveal regional variation or the adequacy of support systems.
Chinese coverage describes a major funding round for an AI inference-services company; the company says it is building token-production capacity around efficiency and reliable delivery.
Also:
Zhidx, Leiphone
More
- The company announced more than 1.3 billion yuan (about $180 million) in Series B funding and says it has raised 2.4 billion yuan (about $340 million) during 2026.
- Its pitch is a managed inference service: focus on a limited set of models, optimize them deeply, and improve both output per unit of compute and effective cluster utilization.
- The company says its systems use heterogeneous inference and KV-cache techniques to raise throughput on domestic accelerators; the article provides no independent benchmark methodology.
- The financing signals investor appetite for Chinese inference infrastructure, where the commercial question is whether operators can deliver stable, high-quality output at competitive unit cost.
- Claims of rapid growth and having crossed the cost line come from company-provided material and should be treated as unverified.
Tencent’s WorkBuddy integration can use authorized WeCom messages, documents, calendars and email, then return generated work into the same collaboration channels, reducing context handoff between people and agents.
More
- The integration targets a real enterprise-agent bottleneck: employees repeatedly reconstruct project context from chats and files before asking an assistant to act.
- Tencent says WorkBuddy can summarize chat history, track decisions and open issues, search WeCom documents and move generated files back into chats or shared storage.
- The report describes sales workflows that connect lead collection, assignment, customer-intent tagging and CRM entry, as well as assistance for customer-facing staff.
- The product’s value depends on permission boundaries and the reliability of extracting decisions from noisy, changing conversations; the article does not describe technical controls in depth.
- The source is a promotional account and gives no independent adoption, accuracy or labor-saving data.
A Chinese AI-generated rap featuring historical and mythical figures passed 100 million views in under two weeks, using familiar cultural icons to voice contemporary anxieties.
More
- The song brings together ten figures from Chinese history and mythology; the source describes its translated title as a phrase meaning someone who has already lived through it.
- Its rapid spread suggests that generative media can gain traction through culturally legible characters and shared social concerns, not only visual novelty.
- The article frames the track as speaking to modern anxieties, though the supplied excerpt does not specify which lyrics or concerns drove the response.
- The view count is reported across platforms and is not broken down by platform, unique viewers or repeat plays.
NetEase's new enterprise system coordinates agents for requirements, design, architecture, coding and testing, aiming to move AI coding from developer assistance to end-to-end delivery.
Also:
InfoQ China
More
- The product uses a specification-driven workflow: an agent clarifies ambiguous requirements, turns them into structured specifications, then routes tasks to design, architecture, development and test agents.
- For legacy systems, the platform can map source code and configuration before changes are planned; NetEase says the system can run in a customer environment and connect to private code and knowledge bases.
- The company reports a 90% first-pass rate on its P0 test cases, but explicitly this is not the same as project-level delivery success.
- NetEase claims 3-to-5-times productivity gains across more than 100 enterprise validations; the source supplies no independent comparison or measurement protocol.
Chinese coverage frames Lula's reelection as preserving Beijing's Latin American foothold while giving Brasilia room to bargain over rare earths between China and the United States.
More
- Lula won 51.8% of the runoff vote, according to the report, while his opponent Flavio Bolsonaro refused to concede.
- The article's central China angle is strategic access: Brazil's rare-earth reserves give it leverage as Washington and Beijing compete for supply and partnerships.
- Lula's continued presidency sustains a government that has overseen deeper ties with China, but the report does not describe a new bilateral deal or policy commitment.
- The outcome is therefore a favorable diplomatic opening for Beijing, not evidence that Brazil will align exclusively with China.
A reported portrait of Chinese women running a bar in a Spanish tourist town traces how migration and family networks create opportunity while leaving owners trapped in long, exhausting work.
More
- The women come from a Zhejiang hometown with a long history of migration to Europe, where kinship and fellow-villager networks have supported work and small-business ownership.
- The bar combines local Spanish food and branding with Chinese touches, illustrating how immigrant businesses adapt to tourist demand without fully disappearing into it.
- The piece connects entrepreneurship to family separation: a daughter reunited with her parents after years apart still feels like a stranger in their household.
- The story is about the human cost and texture of migration, not a representative survey of Chinese businesses in Spain.
Satellite imagery cited by an imagery analyst suggests three possible KJ-3000 production aircraft were at a factory by April, a potentially significant but unconfirmed step in Chinese airborne early warning.
More
- The SCMP report says the imagery could indicate limited production has begun; it does not establish that the aircraft have entered service.
- An airborne early-warning platform can detect and track targets and help coordinate air operations over long distances, making it a force-multiplier rather than simply another aircraft.
- The article links the potential capability to operations across the western Pacific, including a possible Taiwan conflict involving the United States.
- The key uncertainty is whether the apparent production examples are operationally representative aircraft and what sensors, networking and fleet numbers they ultimately carry.
At an industrial fair, three Zhongke Fifth Epoch robots reportedly linked depalletizing, transfer and stacking into one no-human-intervention demonstration, shifting the pitch from a single robot skill to a complete workcell.
More
- The demo required separate robots to hand off material and coordinate timing across multiple stations, a harder systems problem than showing one successful grasp.
- The company says it has deployments in factory material handling and energy-site inspections, plus orders worth several hundred million yuan; the excerpt does not establish how much has been delivered.
- Its embodied model is presented as transferable across robot bodies and able to learn some tasks from fewer examples, claims that would need deployment data to validate.
Kim’s visit to a commercial-vehicle factory in Dandong points to North Korea’s interest in modernizing buses and expanding industrial ties with China.
More
- The report says Kim crossed the newly opened New Yalu River Bridge during a surprise visit to Dandong.
- He toured a commercial vehicle factory, an unusually concrete signal of interest in transport and industrial cooperation.
- The visit was his first trip to China since attending a Beijing military parade the previous September, according to the report.
The short-drama service reportedly had 143 million monthly active users in August, with weekly activity rising sharply in early October as WeChat distribution and increased content supply fueled growth.
More
- QuestMobile data cited by the report puts monthly active users at 143 million in August and weekly active users at 89.68 million in the first week of October, up 44.1% from the first week of September.
- The service benefits from being inside WeChat, where Video Accounts and the wider social graph can feed discovery without requiring a standalone app install.
- The article also cites a rise in paid promotional creative and a mix of live-action, AI-generated and animated short dramas.
- Its scale is evidence of platform distribution and consumer appetite, but the source provides no retention, revenue or profitability figures.
Yangtze Memory's Shanghai listing review remains in questioning while investors cool on chip IPOs, illustrating how quickly China's capital-market appetite for AI-linked hardware can reverse.
More
- Yangtze Memory's proposed 33-billion-yuan raise (roughly $4.6 billion) had been accepted by the Shanghai exchange, but the article says review was still at the inquiry stage after more than two months.
- The report contrasts the muted reception with ChangXin Memory's blockbuster earlier listing, then describes a sharp pullback in semiconductor shares and weak margin financing for another AI supply-chain issuer.
- It argues that the change reflects a narrower IPO window rather than a collapse in underlying memory demand: AI infrastructure still supports DRAM and NAND consumption.
- The article cites Yangtze Memory's rapid revenue and profit improvement, but those figures need to be read alongside cyclicality and the unusually volatile earnings environment.
- For Chinese industrial policy, public-market financing is a key channel for turning state-backed semiconductor capacity into investable assets; a closed window may leave companies more dependent on state and private capital.
A Chinese investigation challenges Wangjingshe’s claim to independent think-tank status, focusing on its paid publicity packages, its past unregistered-association listing and questions about media-service licensing.
More
- The report says the outlet’s website advertised a 200,000-yuan package (about $28,000) combining an overall article, expert commentary and media distribution.
- Wangjingshe denied that a disputed e-commerce commentary was commissioned or paid for, while describing the listed services as ordinary marketing work.
- The article notes that the former China E-commerce Research Center appeared on a 2018 Ministry of Civil Affairs list of suspected illegal social organizations that had not registered with the ministry.
- It also says a newspaper raised questions about whether the outlet’s operating company holds the relevant online news-information service license; the investigation does not establish the regulator’s conclusion.
- The public-interest issue is disclosure: readers may see expert analysis without knowing whether a commercial client funded related content or distribution.
- The controversy illustrates a broader challenge in China’s media environment, where commercial content, research branding and regulated news services can overlap.
Coverage of the iRTE conference argues that conversational AI depends on synchronized speech, reasoning, tools and device response, not just fluent audio generation.
More
- The article places interruption handling, latency and turn-taking alongside speech recognition and model quality as determinants of whether an AI conversation feels natural.
- A proposed fast-and-slow model arrangement keeps real-time dialogue responsive while a separate model handles deeper reasoning and tool calls.
- Chinese firms and research groups discussed customer-service agents, audio processing and AI-enabled devices, reflecting an increasingly integrated local ecosystem.
- The piece is conference coverage and does not provide comparative latency measurements or evidence that the showcased architectures outperform alternatives.
China's overseas short-drama fight is shifting from a challenge to Hollywood into a contest among Chinese platforms, as ByteDance brings a free, ad-supported model into a market dominated by paid apps.
More
- The article says DramaBox and GoodShort were reported to be discussing a strategic merger, while ByteDance's PineDrama had passed 100 million downloads and reached about 7 million daily active users, according to the report.
- For January through September, the article says all ten most-downloaded overseas short-drama apps had Chinese backgrounds, and Chinese apps occupied nine of the top ten in in-app purchase revenue.
- PineDrama's free-with-advertising approach echoes ByteDance's domestic Red Fruit model and could pressure subscription or pay-per-episode competitors to change their economics.
- The piece warns that user acquisition costs already absorb much of the business's margins, while importing a scale-first model risks repeating the domestic industry's growth-with-thin-profits cycle.
Chinese researchers used tree rings and climate data to reconstruct centuries of runoff and project higher flows under future climate scenarios on a river slated to host a major dam.
More
- The research reconstructed runoff from 1325 to 2020 using tree-ring width records from the southeastern Tibetan Plateau.
- The study projects increased runoff and a possible rise in high-flow events under future climate scenarios.
- The finding is relevant to dam planning because hydrology, flood risk and extreme-event assumptions shape the safety case for large infrastructure.
- The report excerpt does not specify the projected increase or establish how the study compares with the dam operator’s design assumptions.