China Desk

China News, Summarized 15 Oct 2026, 20:07 UTC 34 stories 21/23 sources

Today's brief

China's biggest apps just decided an AI agent has to be invited in

151 apps told ByteDance's phone agent no

The vote is in. ByteDance's Doubao phone assistant is an AI that reads the screen and taps through other companies' apps. On 14 October it published the result of its own consultation. 151 apps refused screen automation, 65 of them Tencent's and 23 Alibaba's. WeChat, Taobao, Meituan and DeepSeek are on the list.

Silence meant yes. The protocol, SAEP, is opt-out. Apps had a 30-day window from 14 September, and non-responders defaulted to allowed. PANews counts 155 refusals. Citing QuestMobile, it says only Douyin (ByteDance's own app) and JD.com among the ten biggest apps did not refuse.

A price, not a veto. Doubao tries structured MCP and A2A calls first, meaning interfaces apps open deliberately. Screen control is only the fallback. Refusing the click-through leaves a door: expose an interface, on your terms, revocable. My inference: the big apps are positioning to sell access, not deny it.

Refuse inbound, build outbound. Tencent's WorkBuddy now sits inside WeCom reading messages. Alibaba's Qwen Office plugs into rivals' tools. Invited connectors are fine; uninvited tapping isn't. Nor is this only China. JPMorgan chief executive Jamie Dimon said the bank won't let external agents into customer accounts. Amazon blocked Meta's Muse agent.

Phones are picking other doors. OPPO, whose group shipments fell over 20% in the third quarter per Counterpoint, is plugging Manus and ChatGPT into its assistant. StepFun's agent-first phone goes on sale tomorrow. Apple's route runs through the regulator.

Go deeper on this section: ClaudeChatGPT

Beijing waves through Apple's easy features

Cook announced it in Shanghai. Apple executive chairman Tim Cook said Apple Intelligence and Alibaba's Qwen models are approved. A beta is due late October with live translation, writing tools and photo clean-up. Siri AI is not included.

Notice who carried it. The state-backed outlet The Paper ran the interview, and Reuters relayed it. Commerce Minister Wang Wentao met Cook the same day, IT Home reports.

THE TELL: Siri AI is the part that would act across apps, the layer those 151 refusals just fenced. Yesterday I noted it had skipped Brussels and Beijing, and Beijing's half is now confirmed. My inference: content tools clear first, delegated action last.

A hotline is being drafted. SCMP's sources say US and Chinese technical teams are finalising an AI-incident framework. They hope to announce it when Xi and Donald Trump meet at APEC in Shenzhen next month.

Go deeper on this section: ClaudeChatGPT

Exports surged and Brussels hardened

Exports rose 36.3%. September's jump lifted the monthly surplus to $133.5bn and the nine-month total to $938bn. THE NUMBER: semiconductor exports are up 108.1% for January–September and autos 54.3%, AP reports.

Two framings. AP leads on mounting pressure to rebalance. Premier Li Qiang told the Canton Fair that China is as big a buyer as a seller, and that selling benefits its neighbours, per Sinocism. One is a complaint, the other a rebuttal.

Brussels read the hybrid deal as proof pressure works. SCMP reports that EU planners now plan to keep talking while tightening. My read: a hybrid cap and a rebate cut are small beside a 36% month.

Go deeper on this section: ClaudeChatGPT

Zhipu's cyber bet, and idle GPUs

Zhipu hired credibility. Yuki Chen is a five-time top researcher on Microsoft's security-response leaderboard and former head of 360's vulnerability institute. He joined this week. InfoQ China says Microsoft's latest patch round credited 23 bugs to "Yuki Chen using GLM from z.ai". One scored 9.8 of 10, though InfoQ and Zhidx name different software. Zhidx says about 100 security firms now plug GLM in.

I was half wrong yesterday. I said nobody aimed at Zhipu in Anthropic's worm paper, and that still holds. But InfoQ's recap says NIST's evaluation centre rated GLM-5.3 the strongest open-weight cyber model on 17 September, roughly four months behind US frontier. Anthropic's 29 September report had it completing 50 of 410 exploit attempts against 56 for its Mythos Preview. Evaluators keep testing Zhipu because cyber is what it sells. That is the incident class a Shenzhen hotline exists for.

Utilisation is the shortage. One lab scheduling specialist told Leiphone that industry GPU utilisation runs 25–27%, against 40% for good infrastructure teams. Top infra engineers now earn 7–8m yuan (~$1.0–1.1m), though one founder says only 5–10% capture that premium. InfoQ China adds that past a few thousand cards, communication and fault-finding decide real compute.

Ant's Ling-3.1-flash attacks the other side. It is open weights at 560B total and 25B active parameters, with 1M context. Quantising routed experts to MXFP4 cuts weights from about 534GB to 292GB and serving from eight Hopper cards to four. Artificial Analysis scores it 41, at about 210 tokens a second and roughly $0.19 per million tokens, InfoQ China reports. Jensen Huang told podcaster Lex Fridman that scarcity forces Chinese labs toward efficient architectures. Ling is the receipt.

DeepSeek's API went dark for five hours. It was down from 8:11 to 13:04 Beijing time. The notice promises credits equal to seven days' spend and to strip out "single-point dependencies", per V2EX. That lands in the week of a reported 100bn yuan (~$14bn) round.

Go deeper on this section: ClaudeChatGPT

Harbin's informer, Luckin's algorithm

Twenty students were punished for VPNs. China Digital Times reports Harbin Engineering University disciplined seven undergraduates and 13 graduate students after a two-day sweep. Unverified rumour says a classmate informed to improve her graduate-admission ranking. She reportedly admitted using a VPN herself for papers and GPT. Two of five archived commentaries were later deleted.

My read: after yesterday's Pei Yi and Kevin Tsai cases, Harbin shows the incentive underneath. Informing pays, and the rule enforced is one academic research can't obey.

Luckin's rota is an algorithm. Initium's podcast says AI schedules staff from sales, orders and hours. Managers say it keeps compressing headcount. Unpaid overtime covers what the system can't see: training, cleaning, stock counts.

Go deeper on this section: ClaudeChatGPT

Threads we are pulling

  • Chabaidao went from halt to index exit. Following yesterday's SFC freeze, Hang Seng Indexes will drop the tea chain after the close on 22 October at a nominal HK$0.0001, per 36Kr's digest. Nobody at the company has been accused.
  • The bubble warning got a price. After yesterday's STAR 50 slide and the chip-equipment chairman's caution, the ChiNext index fell 3.1% on 14 October, with AI-compute names down 5–7%. Robot maker Unitree opened at 1,100 yuan on 19 August and now trades near 420. Its market value has fallen from about 444.9bn yuan (~$63bn) to under 170bn (~$24bn), TMTPost reports.
  • Orbbec shows what the froth hid. The 3D-sensor maker trades at 278 times earnings. Insiders sold about 780m yuan (~$110m) in June and July. Its filing forecasts 2026 related-party sales to Ant Group down 68% to 80m yuan (~$11m), TMTPost reports.
  • The Claude ban thread. A V2EX developer says a year-old Max account was banned and calls the switch to Codex withdrawal. Sentiment, not data.

Go deeper on this section: ClaudeChatGPT

The river · 34

Twenty Harbin students disciplined for unauthorized VPN use, sparking a debate over informers

# China Digital Times 14h

A university disciplined 20 students for unauthorized VPN use; Chinese online discussion focused on alleged peer reporting, academic rivalry and the weaponization of reporting systems.

More
  • The university says seven undergraduates and 13 graduate students were disciplined after a two-day enforcement period in late September.
  • The more consequential controversy is the reported claim that a student informed on a rival to gain an admissions advantage; the account is circulating online and is not established by the university notice.
  • Archived commentary links the case to the reliance of researchers on overseas academic resources and to anxiety about reporting systems being used for personal retaliation.
  • The episode shows how enforcement of internet controls can intersect with competitive academic admissions and everyday interpersonal trust.
  • Discussion also connects the case to a perceived wider chilling effect on online speech, though the source provides commentary rather than evidence of prevalence.
  • This is a rare view into how censorship rules are experienced in academic life, where access to global research tools can be practically important.
Open at China Digital Times Ask Claude ChatGPT

Huawei says Kirin 9050 uses stacked logic to improve chip efficiency

# Leiphone 17h

Huawei says its Kirin 9050 Pro is the first mass-produced Kirin chip using logic folding, a 3D layout linked by more than 50 million hybrid bonds to shorten data paths.

Also: Zhidx

More
  • Huawei reports a 1.5-micrometer hybrid-bond pitch and more than five million effective signal connections, using vertical links to place related logic across layers.
  • The company says the approach raises transistor density by 53.5% and improves CPU, GPU and NPU performance and power efficiency; these figures are vendor claims and are not a process-node equivalence.
  • A key architectural target is shorter paths between CPU cores and cache, which can reduce clock-distribution overhead and create room to raise frequency or lower voltage.
  • Huawei also says a 30B MoE model can run with less than 2.5GB of additional memory by keeping active experts in DRAM and inactive experts in flash.
  • This is a system design involving chip layout, bonding, memory hierarchy and model scheduling, not merely a new processor block.
  • The mass-production claim matters because it moves the design from a research concept into a shipping flagship product, though independent teardown and benchmarking remain important.
Open at Leiphone Ask Claude ChatGPT

China and US discuss an AI incident channel ahead of a leaders' summit

# SCMP 10h

Chinese coverage leads with regular bilateral AI talks and plans for an incident hotline, while the immediate summit context is the effort to formalize a process before Xi-Trump meetings.

More
  • The reported goal is a basic framework for addressing AI incidents before the leaders meet at the APEC gathering in Shenzhen.
  • An incident channel could help manage risks such as agent failures, cyber operations or other fast-moving events, but the source gives no agreed protocol or scope.
  • A hotline is only useful if officials can exchange timely, actionable information and have clear procedures for escalation; those institutional details remain unsettled.
  • The effort sits alongside strategic competition, making it a narrow risk-management measure rather than evidence of broader trust.
  • The report does not establish that a final agreement has been reached.
Open at SCMP Ask Claude ChatGPT

EU leaders consider a trade tool to eject Chinese firms from the market

# SCMP 3h

Chinese coverage leads with Brussels' discussion of a new trade weapon that could rapidly exclude Chinese companies; the reported proposal has broad backing among EU member states.

More
  • The initiative follows a French-German paper and could give the bloc a faster way to respond to economic coercion or other perceived threats, though the source says the instrument is still being developed.
  • The framing matters: this is not presented merely as another tariff or investment-screening rule, but as a tool capable of removing firms from the EU market.
  • A mechanism this broad would raise major questions about triggers, due process, proportionality and the treatment of companies with extensive European operations.
  • For Chinese companies, the risk is that market access becomes conditional on geopolitical assessments as well as ordinary competition and compliance standards.
  • The reporting describes political support and Commission openness, not an adopted law or an operative enforcement power.
  • Its practical force will depend on the final legal design and whether leaders agree on a threshold for using it.
Open at SCMP Ask Claude ChatGPT

China took half of Brazil's oil exports as Iran war disrupted Gulf supply

# SCMP 1h

China's trade with Brazil hit $143 billion in the first nine months as Brazilian oil sales surged after war-related disruption to Gulf supply redirected energy flows.

More
  • The Brazil-China Business Council says Brazilian exports to China rose 12% to $84.7 billion, while imports from China rose 9.9% to $58.6 billion in January-September.
  • The reported oil shift illustrates how a disruption around the Strait of Hormuz can rapidly strengthen China's role as a buyer of alternative crude supplies.
  • The trade relationship is strategically asymmetric: Brazil ran a $26.1 billion surplus with China over the period, while Chinese demand gives Brazilian commodity exporters a major market.
  • A sustained change would depend on the duration of Gulf disruption, shipping and insurance conditions, and whether Brazilian producers can maintain export volumes.
  • The source cites council analysis of government data; it does not establish how much of the oil shift will persist after the conflict or supply shock eases.
Open at SCMP Ask Claude ChatGPT

151 Chinese apps refuse ByteDance's phone agent permission to operate their screens

# Leiphone 9h

Chinese coverage centers on 151 apps, including WeChat, Taobao and Meituan, refusing GUI automation by ByteDance's phone assistant; users can still operate the apps manually.

Also: 36Kr

More
  • The refusal is about an agent visually reading screens and simulating taps, not ordinary installation or human use of the apps.
  • The episode exposes a basic platform-governance conflict: cross-app agents need broad interface access, while app operators retain control over automation on their services.
  • If major social, commerce and payments apps opt out, a phone agent may be unable to complete the very multi-app tasks that justify its existence.
  • The episode also shows that a permission protocol can shift the decision to individual developers rather than creating a universal automation interface.
  • The same roundup mentions a Tencent short-drama product and debate over short-video consumption, but those items are separate developments rather than evidence about GUI access.
Open at Leiphone Ask Claude ChatGPT

China opens a long-delayed bridge to deepen trade with North Korea

# SCMP 10h

Premier Li Qiang used the opening of the New Yalu River Bridge to frame the link as a way to expand trade and people-to-people exchanges with North Korea.

More
  • The bridge opening is a concrete infrastructure signal after a long delay, though the report does not say when or at what capacity regular traffic will begin.
  • Beijing's public framing emphasizes friendship, trade and personnel exchanges, presenting connectivity as a tool for stabilizing and deepening the bilateral relationship.
  • The significance will depend on actual use: border policy, sanctions enforcement and North Korea's domestic economic conditions will shape whether the bridge carries meaningful trade.
Open at SCMP Ask Claude ChatGPT

Apple says China testing of Apple Intelligence with Qwen starts this month

# Policy Leiphone 10h

Tim Cook said Apple Intelligence features using Alibaba's Qwen model have regulatory approval and are due to enter testing in China later in October; Siri AI is excluded.

More
  • The initial features include live translation, writing tools and photo cleanup, indicating a staged rather than full-feature launch.
  • The arrangement shows how Apple's China AI offering depends on local model integration and regulatory approval, not only on Apple's global product roadmap.
  • The source reports testing plans and approval, not general availability or the scope of the final China release.
Open at Leiphone Ask Claude ChatGPT

Top security researcher Yuki Chen joins Zhipu

# AI Zhidx 17h

Yuki Chen joined Zhipu after Microsoft credited 23 vulnerabilities he found with GLM assistance, including a severe OpenSSH flaw, according to the company and researcher.

Also: InfoQ China

More
  • The report says Microsoft's security response center reviewed and confirmed the credited reports before public acknowledgment.
  • Zhipu says roughly 100 security firms have integrated GLM into products or workflows, but the source gives no revenue figures or independent customer verification.
  • The hire suggests Chinese model companies are building security research into their product and commercialization strategies, not treating it only as a benchmark capability.
Open at Zhidx Ask Claude ChatGPT

StepFun opens Step 5 Preview weights for long-running agents

# AI InfoQ China 0m

StepFun released BF16 and FP8 weights for a 600B-parameter sparse MoE model with 27B active parameters and a one-million-token context, targeting long-horizon agent work.

Also: IT Home, Zhidx

More
  • The model uses a 92-layer Transformer, alternating gated sliding-window attention with sparse global attention; the team says grouping adjacent tokens cuts indexing and selection overhead by eightfold, not total inference cost.
  • Its design targets software tasks that span file edits, tool calls, test runs and corrections, where a model must preserve original constraints and intermediate state across a long execution.
  • StepFun says context compression is trained as part of the agent policy, so the model learns to summarize recoverable task state as its context budget fills; reliability after repeated compression remains to be demonstrated.
  • The reported million-scale training environments use verifiable tasks, but environment diversity and resistance to benchmark gaming matter more than the headline count.
  • The 27B active-parameter figure does not make deployment equivalent to a 27B dense model: all expert weights still need to be stored or otherwise managed.
  • Open weights and BF16/FP8 variants let enterprises test the system on their own repositories and workflows, where completion rate, runtime, resource use and human handoffs are the meaningful measures.
Open at InfoQ China Ask Claude ChatGPT

Shanghai AI Lab says its model helped prove a longstanding mathematics conjecture

# AI Leiphone 12h

Shanghai AI Lab says researchers used its S2 model in a human-led proof of the three-dimensional Ghys conjecture, later checked with formal verification.

More
  • The account describes the model producing long proof branches while mathematicians selected promising directions and corrected the exploration over repeated rounds.
  • The report says the proof graph expanded from 15 to 389 nodes and the team completed seven rounds of continued reasoning, illustrating a human-and-model research loop rather than autonomous theorem proving.
  • A separate formalization framework was used to verify the proof, an important distinction between generating plausible mathematics and checking correctness.
  • The paper and formal code are said to be public, but independent mathematical scrutiny remains the decisive test of the claimed result.
  • If confirmed, the work would be notable less as replacement of mathematicians than as a system for exploring many proof paths and handing experts a richer set of candidates.
Open at Leiphone Ask Claude ChatGPT

A Chinese robotics team tackles the gap between walking and working

# AI GeekPark 9h

YuanCe Future demonstrated a whole-body control approach that coordinates balance, reaching and movement, including stepping onto a platform to reach an object.

More
  • The system separates a high-frequency motion-control model from a higher-level task-planning model, assigning balance and execution to one layer and task decomposition to another.
  • The company reports reducing hand-end tracking error to around two centimeters and extending control beyond flat floors; these are company test claims.
  • The architectural point is that mobile manipulation requires the legs, torso and arms to operate as one control problem rather than walking to a spot and freezing before manipulation.
  • Whether the approach generalizes across robot bodies, surfaces and unscripted tasks is not established by the demonstrations described.
Open at GeekPark Ask Claude ChatGPT

Chinese interconnect-chip maker Zonsen unveils a broader product platform

# Analysis Zhidx 8h

Wuxi-based Zonsen presented products spanning storage, networking, I/O expansion and AI-cluster interconnect, positioning itself against established foreign suppliers.

More
  • The company's portfolio includes PCIe switching, network adapters and storage products; it says its PCIe switch can substitute for a Broadcom product in tested configurations.
  • The technical argument is that cluster performance depends on moving data among compute, memory and network devices, not only on accelerator peak throughput.
  • Zonsen reports more than 800,000 cumulative shipments and partnerships with Chinese server and software firms, but the article is company-centered and gives limited independent performance evidence.
Open at Zhidx Ask Claude ChatGPT

Returning scientist says China is now a leading center for embryo research

# Analysis SCMP 6h

Stem-cell researcher Wu Jun returned from the United States to lead a lab at Westlake University, arguing that many top embryo researchers now work in China.

More
  • The profile places the move within a broader contest for scientific talent and research capacity, rather than treating it as an individual career change alone.
  • Wu is known for work on stem-cell-derived embryo models, an area that combines frontier biology with difficult ethical and regulatory questions.
  • The article offers the scientist's assessment of the field, not a comparative dataset establishing China's global rank.
Open at SCMP Ask Claude ChatGPT

China's September trade, prices and credit data take center stage

# Signal Sinocism 21h

Sinocism's China briefing highlights September trade, price and credit data, Li Qiang's remarks at the Canton Fair and EU-China issues; it reports exports up 36.3% year over year.

More
  • The newsletter says imports rose 31% and the monthly trade surplus reached $133.53 billion, the second largest on record.
  • Li Qiang's cited framing presents China as both a major exporter and buyer, emphasizing benefits to trading partners rather than only China's surplus.
  • The excerpt signals coverage of price and credit data but does not include enough detail to assess those indicators.
Open at Sinocism Ask Claude ChatGPT

Chinese drugmaker puts its information department under the president in an AI push

# Analysis 36Kr 16h

Drugmaker Simcere's president has taken direct oversight of IT and says managers must use AI, with technical staff embedded in business teams and drug-research workflows.

More
  • The company is applying AI to early target discovery and molecule design, while also exploring use in manufacturing, clinical work, marketing and operations.
  • Simcere says innovative medicines now account for about 85% of first-half revenue, giving the AI effort a strategic role in a shift away from its older generics business.
  • The executive argues that early-stage discovery is a better initial fit for AI than later regulated drug development, where clinical validation and accountability remain human responsibilities.
  • This is an organizational change as much as a technology project: executive sponsorship is being used to overcome weak adoption inside an established pharmaceutical company.
  • The report provides no measured productivity gain or evidence that AI has yet produced a commercial drug candidate.
Open at 36Kr Ask Claude ChatGPT

Kunlun Tech open-sources a real-time video-model game for one consumer GPU

# AI Leiphone 14h

Kunlun Tech released Matrix-Game 4.0, a game system that pairs an explicit engine-managed world with a video model rendering visuals in real time.

More
  • The company reports 480p at 60 frames per second on one RTX 5090, with under 150 milliseconds from player input to visual response.
  • A game engine stores rules and world state, while the video model generates the rendered scene; this separates deterministic game logic from visual generation.
  • The design addresses a core weakness of video-generated worlds: models can drift on object identity, off-screen state and exact game rules over long interactions.
  • The project is open source, but the reported performance is tied to a high-end GPU and does not establish that generated worlds can support complex commercial games.
Open at Leiphone Ask Claude ChatGPT

Why more domestic GPUs can still mean less usable compute

# Analysis InfoQ China 42m

A Chinese infrastructure profile argues that domestic accelerators have moved the bottleneck from card availability to system-level efficiency across networking, scheduling and operations.

More
  • The article describes model migration as much more than matching framework APIs: operators, communication libraries, memory management, monitoring and failure diagnosis may all need retuning.
  • Small timing mismatches between compute and communication can leave accelerators idle; at thousands of devices and many training iterations, modest per-step delays compound into substantial lost capacity.
  • A mixed fleet of accelerator generations and vendors creates a second challenge: companies must reconcile different software stacks and telemetry rather than treating all cards as interchangeable.
  • The strategic shift is from nominal peak FLOPS toward delivered cluster throughput and reliability, making interconnect and orchestration capabilities central to domestic compute competitiveness.
Open at InfoQ China Ask Claude ChatGPT

Agent demand renews China's push for on-device AI

# Analysis TMTPost 9h

At a Chinese industry conference, chip and model executives argued that agent workloads could accelerate edge AI, while researchers highlighted interface governance and deployment limits.

More
  • A speaker cited a two-to-three-order-of-magnitude increase in token consumption from agents, while the chipmaker reported more than 150 product design wins for its M50.
  • The conference discussion treated edge AI as a shift in compute location that also redistributes power, risk and responsibility between device makers, app platforms and users.
  • Speakers argued that local inference still needs cloud coordination for many complex tasks; data access, workflow integration and reliability remain obstacles.
  • Predictions that edge models will handle most everyday work within one to two years are executive forecasts, not established adoption timelines.
Open at TMTPost Ask Claude ChatGPT

Wang Yi calls for Asian cooperation on security, trade and governance

# Policy Jiemian 5h

At the CICA summit, China's foreign minister urged regional cooperation on security, trade corridors, supply chains and digital governance, while opposing bloc confrontation.

More
  • The speech links regional development to stable transport routes and access to critical materials, energy and food, not only to diplomatic security language.
  • Wang called for coordination on the Middle East, Afghanistan and key shipping routes, including Hormuz and the Red Sea.
  • The governance agenda includes stronger UN authority and greater regional capacity in digital governance, disaster response and public health.
Open at Jiemian Ask Claude ChatGPT

OPPO's wearable recorder aims to turn conversations into work memory

# Culture GeekPark 2h

OPPO's AI clip records continuously, turns speech into searchable text and groups discussions into project timelines, positioning the phone maker against standalone recorder startups.

More
  • The product's proposed workflow is more than transcription: it links separate conversations into a topic timeline and extracts tasks and decisions for later review.
  • OPPO says it keeps text rather than original audio and signals recording with a visible light, tradeoffs intended to address the social cost of wearing a microphone all day.
  • The reporting is based on a product test and company claims; sustained accuracy, consent in shared spaces and the usefulness of its behavioral insights remain unproven.
Open at GeekPark Ask Claude ChatGPT

China-led team reports an 83% success rate in an earthquake-signal study

# Signal SCMP 8h

A China-led research team says ring-shaped seismic, magnetic and ionospheric disturbances preceded major earthquakes in its study, a finding that needs independent validation.

More
  • The researchers call the reported patterns seismo-ripples and say their monitoring began in 2022; the source does not provide enough methodological detail to assess the claimed success rate.
  • Earthquake prediction is a notoriously difficult scientific problem, so prospective replication and clear false-alarm rates matter more than a retrospective headline percentage.
Open at SCMP Ask Claude ChatGPT

Ant Group's Ling model handles a newsroom's multi-step office tasks in a test

# AI Zhidx 14h

A newsroom test of Ling-3.1-Flash had it analyze financial reports, organize work materials and build interactive prototypes, alongside the model's open-weight release.

Also: InfoQ China

More
  • The test asked the model to retrieve and compare financial data, create a brief and editable table, and generate an interactive dashboard.
  • It also produced a task workbench from simulated email and chat records, organizing 33 pending items by status and priority.
  • The article's demonstrations show a workflow from requirements through code to a usable artifact, but they are publisher-run tests rather than independent evaluations.
  • This is the same Ling-3.1-Flash release covered by the model's open-source announcement; the additional angle is applied task performance.
Open at Zhidx Ask Claude ChatGPT

Deep Robotics returns to loss as its IPO filing leans on robot dogs

# Analysis TMTPost 9h

Deep Robotics reported a first-half loss after a profitable 2025, with quadruped robots still driving sales and humanoid deliveries remaining minimal.

More
  • The filing data cited by the article show revenue of 259 million yuan (about $36 million) in the first half, but a net loss of 8.8 million yuan (about $1.2 million).
  • The company says lower-margin bulk deliveries and higher R&D spending weighed on results; its industrial quadruped Shining X average selling price also fell.
  • The contrast with a humanoid-focused market narrative matters: the company's current revenue base is specialized inspection robots, not humanoids.
  • Customer concentration and the possibility that a major buyer develops competing products are material risks to the IPO case.
Open at TMTPost Ask Claude ChatGPT

OPPO brings Manus agents into its phone software as handset sales fall

# Analysis TMTPost 9h

OPPO plans to integrate Manus multi-step agents into its AI Mind Pilot, even as third-party estimates show the phone maker's shipments declining amid a weak market.

More
  • The article contrasts OPPO's choice to add an external agent platform with competitors pursuing tighter model-to-operating-system integration.
  • Reported estimates put OPPO's China shipments down 17.6% year over year in the quarter, while its global family of brands fell more than 20%; these are third-party figures.
  • Higher-end sales growth has not yet been shown to offset overall volume pressure, and the company has not disclosed comparable phone-profit data.
  • The strategic question is whether an agent partnership can create a distinctive reason to buy a high-end OPPO phone, rather than simply adding another AI feature.
Open at TMTPost Ask Claude ChatGPT

SK Hynix seeks US approval for more equipment at its China factory

# Policy IT Home 5h

SK Hynix is reportedly seeking permission to install more US-made tools in its Dalian NAND facility while weighing new US manufacturing capacity.

More
  • The proposed arrangement illustrates how export licensing links a foreign semiconductor company's China production plans to its investment decisions in the United States.
  • The source says the company has applied for licenses for equipment in an unfinished area of the Dalian site, while SK Hynix says installation is being discussed with Washington and is not a request to expand capacity.
  • The strategic tension is between preserving existing China-based production and satisfying US pressure to build more domestic capacity.
  • Whether the plan proceeds depends on US authorization; the report does not say that the license has been granted.
Open at IT Home Ask Claude ChatGPT

Phanes AI and Tianlong launch a wearable system for robot-training data

# AI Leiphone 12h

Chinese startups Phanes AI and Tianlong introduced a wearable capture system combining first-person video, full-body motion tracking and tactile gloves.

More
  • The system combines a four-camera headband, wrist devices, waist and leg trackers, and a tactile glove to capture body movement and contact forces alongside visual input.
  • Its premise is that first-person video alone misses important state, including how the body moves outside the camera view and how much force the hands apply.
  • A processing platform converts recordings into segmented, annotated data and maps human movements to robot bodies; the value will depend on transfer quality and usable task diversity.
  • The product is aimed at a real bottleneck in robot learning: whether richer data per hour improves manipulation more than simply collecting longer recordings.
Open at Leiphone Ask Claude ChatGPT

China and Russia discuss US ties ahead of a possible three-leader meeting

# Policy SCMP 11h

Wang Yi and Sergey Lavrov discussed US relations and international conflicts, prompting speculation about a possible Xi-Trump-Putin meeting at the next APEC summit.

More
  • The source says Moscow described the ministers' discussion as covering US relations as well as the Ukraine and Iran crises.
  • A three-way leaders' meeting is presented as a possibility, not a scheduled event or confirmed diplomatic plan.
  • The item reflects the way China-Russia coordination is being linked to the summit agenda, but offers little evidence of substantive policy alignment.
Open at SCMP Ask Claude ChatGPT

An AI service matched six strangers for a night out

# Culture 36Kr 14h

A Chinese startup uses AI behind the scenes to match paying participants into small-group dinners and coordinate venues, while leaving the actual conversation to people.

More
  • The operator says its agents handle matching, substitutions, venue coordination and custom discussion cards across events in 14 cities.
  • The product avoids AI-mediated conversation: its bet is that software can reduce the operational cost of organizing social contact without replacing the human interaction itself.
  • The company reports more than 40,000 paying users and angel funding, but the article does not provide retention, unit economics or repeat-booking rates.
  • The model tests whether AI can scale a service business whose value comes from bringing people together rather than generating content.
Open at 36Kr Ask Claude ChatGPT

Orbbec's robotics revenue remains opaque as insiders sell shares

# Signal TMTPost 9h

An analysis questions how much revenue Orbbec actually earns from robot vision, noting that public filings do not break out robotics sales despite the company's market-position claims.

More
  • The source says Orbbec's filings report revenue by product category, not end market, making it difficult to isolate robotics demand from other 3D-vision uses.
  • It reports that the company's largest customer orders fell sharply and that a substantial portion of recent profit came from non-recurring gains.
  • Insider share sales and a new Hong Kong listing plan raise governance and valuation questions, but the article's strongly critical framing is not a substitute for the company's response.
Open at TMTPost Ask Claude ChatGPT

Tmall says a record 1 million brands are joining Double 11

# Signal Leiphone 9h

Alibaba's Tmall says 1 million brands and 72 million products will participate in this year's Double 11 shopping festival, both platform records.

More
  • The platform is presenting AI and flagship stores as central to its sales strategy, claiming AI can deepen its understanding of consumer demand.
  • The figures are promotional participation counts, not evidence of sales growth or stronger household spending.
Open at Leiphone Ask Claude ChatGPT

Why Chinese commodity futures can swing far beyond current supply and demand

# Signal TMTPost 10h

A market analysis uses lithium and soda ash to show how futures prices can move on expectations, delivery constraints and leveraged trading rather than current demand alone.

More
  • The article reports the lithium contract falling below 100,000 yuan per ton while spot prices remained higher, illustrating the difference between delivery today and a standardized future contract.
  • For soda ash, the analysis distinguishes total inventories from immediately deliverable exchange-grade stock, which can amplify near-term price spreads.
  • The source cautions that a change in the inventory sample can break comparisons, a reminder that market data series need consistent definitions.
  • These examples are useful for understanding Chinese commodity pricing, but the article does not establish a forecast for future prices.
Open at TMTPost Ask Claude ChatGPT

Jensen Huang says Nvidia should work with Chinese AI firms on safety

# Policy IT Home 6h

Nvidia CEO Jensen Huang said dialogue with Chinese AI companies is needed to monitor risks from advanced AI, amid discussion of a bilateral incident channel.

More
  • Huang frames cooperation as risk monitoring rather than a claim that the two countries have aligned on safety rules.
  • The source says the US Treasury secretary had proposed a bilateral channel covering potential agent failures, cybersecurity and biological risks.
  • A recurring dialogue mechanism would be meaningful only if it has agreed reporting practices and survives broader technology and security tensions.
Open at IT Home Ask Claude ChatGPT

AI infrastructure engineers see a sharp rise in pay as compute costs bite

# Culture Leiphone 12h

Chinese recruiters and executives say infrastructure engineers have become more valuable as model firms focus on GPU utilization, inference costs and cluster efficiency.

More
  • The article gives recruiter anecdotes of salaries rising from tens of thousands of yuan to around 1 million yuan annually, with some executive estimates reaching 7 million to 8 million yuan (roughly $1.0M-$1.1M).
  • One interviewee estimates typical GPU utilization at 25% to 27% and says strong infrastructure teams can raise it to 40%; the figures are industry estimates, not a standardized survey.
  • The underlying shift is from rewarding benchmark gains toward rewarding measurable savings in accelerator hours, networking and serving costs.
  • The pay surge is selective: the article says compensation depends on specialized experience in kernels, scheduling and distributed inference, not simply an infrastructure job title.
Open at Leiphone Ask Claude ChatGPT
Source ledger
ChinAI 403 0/0 HTTPError: 403 Client Error: Forbidden for url: https://chinai.substack.com/feed
ChinaTalk 200 0/20
Pekingnology 200 0/20
Ginger River Review 200 0/20
Sinocism 200 1/20
QbitAI ERR 0/0 ReadTimeout: HTTPSConnectionPool(host='www.qbitai.com', port=443): Read timed out. (read timeout=30)
36Kr 200 7/30
Solidot 200 10/20
Leiphone 200 16/20
GeekPark 200 3/30
IT Home 200 12/60 +29 over cap
SCMP 200 10/50 +10 over cap
Initium 200 3/15
Jiemian 200 8/30 +14 over cap
China Digital Times 200 1/15
The Wire China 200 0/50
Techmeme 200 2/15
Hacker News 200 0/20
V2EX 200 8/50 +35 over cap
Zhidx 200 8/20
InfoQ China 200 8/20 +12 over cap
New Wisdom 200 8/15 +2 over cap
TMTPost 200 8/19 +11 over cap