China Desk — 23 Sep 2026

DeepSeek opens the sandbox infrastructure behind its agent training

# Zhidx

DeepSeek’s DSec platform reportedly serves three million sandboxes a day, peaks above 380,000 concurrent environments and can launch 32,000 sandboxes for one training task.

Agent reinforcement learning needs real execution environments for code, tools and software, each with state that must be isolated, reset and preserved between rollouts; ordinary container-per-task designs waste memory while agents wait.
DSec supports function calls, containers, Firecracker microVMs and full virtual machines behind one Python interface, allowing the training framework to choose isolation without rewriting its control logic.
The platform separates base images, workspaces and tools into composable read-only layers and loads only the image data a sandbox actually reads, reducing burst deployment time and disk writes.
DeepSeek reports a production unit with about 160 CPU nodes, 30,000 cores and 250 terabytes of memory, showing that agent training environments can become a major infrastructure tier beside GPU training.
From version 4.1 onward, the company reportedly moved rollout execution away from training GPUs, reducing interruption and allowing environment capacity to scale independently.
The strategic lesson is that reliable agent training depends as much on operating-system and cluster engineering as on model architecture; without cheap, reproducible environments, reinforcement learning cannot scale.

Open at Zhidx Ask Claude ChatGPT

Zhipu open-sources ZCode after a repository snapshot controversy

# InfoQ China

Zhipu says it removed a repository snapshot upload path from ZCode after researchers found encrypted code histories could reach cloud storage, then released the product source.

The investigation found local snapshots containing Git history, LFS data and reflogs; a small test repository was uploaded while a larger 313 MB archive remained pending locally.
Zhipu said the behavior came from repository indexing and Wiki generation, apologized, deleted the relevant objects and commissioned third-party checks.
The response is notable because open-sourcing the desktop, web, backend, CLI and runtime components exposes the product to outside audit instead of relying only on assurances.
The architectural issue is broader than one upload path: the agent can use shells, plugins, MCP servers, browsers and remote workspaces, and the shared runtime lacks an operating-system sandbox by default.
Encryption did not eliminate the trust problem because the server controlled key wrapping and the client sent ciphertext directly to cloud storage, limiting user visibility into policy enforcement.
Open source can verify that the reported path was removed, but it does not make the agent local-only; credentials, browser sessions and tool permissions remain part of the attack surface.
The event is a useful enterprise lesson: a coding agent's data boundary is defined by its entire execution graph, not by its headline repository feature.

Open at InfoQ China Ask Claude ChatGPT

China moves its industry minister to lead Anhui

# SCMP

China appointed Industry Minister Li Lecheng as Communist Party secretary of Anhui, a major semiconductor and manufacturing province, in a significant personnel move.

Li has led the Ministry of Industry and Information Technology since April 2025, so moving him to Anhui links provincial leadership directly to industrial and semiconductor priorities.
The appointment replaces Liang Yanshun, who had held the Anhui Party leadership since June 2024, indicating a rapid reshuffle at a strategically important provincial post.
Anhui hosts major semiconductor, automotive and advanced-manufacturing activity, making the move more consequential than an ordinary provincial rotation.
The personnel choice may signal stronger central coordination of industrial policy, local implementation and supply-chain resilience.
It also shows how senior technocrats can be redeployed from national ministries to provinces where policy execution and industrial upgrading are tightly coupled.
The practical test will be whether Anhui receives clearer capital, land, talent and procurement support, not merely a new political sponsor.

Open at SCMP Ask Claude ChatGPT

Huawei links 4,096 AI chips into one supernode

# InfoQ China

Huawei's Ascend 960 supernode uses near-packaged optics and unified memory addressing to connect 4,096 accelerators for trillion-parameter training and inference.

Huawei says 5,500 optical engines replace 48,000 800G optical modules, cutting power by more than 550 kilowatts; these are vendor claims, not independent measurements.
The supernode exposes multiple physical servers as one logical machine through a unified interconnect and memory model, attacking communication overhead that Huawei says can consume over 40 percent of training time.
Near-packaged optics shortens high-speed electrical paths, but forces tighter optical, thermal, mechanical, and reliability co-design near the accelerator package.
The system uses liquid cooling and is rated at up to 8 exaflops of FP8 compute, placing networking and thermal architecture alongside accelerator throughput as the competitive unit.
Huawei also positions a 4,096-node Kunpeng system for agent sandboxes, where startup latency and density become as important as model inference.
The strategic signal is China's push to compete at system architecture level when access to leading foreign accelerators is constrained.

Open at InfoQ China Ask Claude ChatGPT

Alibaba's Zhenwu V900 targets the system around the AI chip

# GeekPark

Alibaba's T-Head unveiled the Zhenwu V900 with 216GB of memory and 1,200GB/s interconnect bandwidth, alongside networking, storage, and CPU components for large AI systems.

The chip is designed for both training and inference, with native FP8 and FP4 support and a claimed threefold performance gain over the prior M890.
Alibaba's ICN Switch provides unified addressing and memory semantics so hundreds or thousands of accelerators can operate as one high-bandwidth system.
The product family includes smart NICs, SSD controllers, and CPUs because mixture-of-experts models shift bottlenecks among arithmetic, memory capacity, and inter-chip communication.
The strategic shift is from selling an accelerator to controlling the compute, storage, network, and software interfaces around it.
The V900 is planned for mass production in 2027, so customer deployment and independent performance remain future questions.

Open at GeekPark Ask Claude ChatGPT

Qiyuan puts consumer humanoid robots on sale in China

# Leiphone

Qiyuan Robotics began selling two consumer humanoid robots at 19,999 yuan (roughly $2,800), claiming an industry first for mass-market delivery rather than demonstrations.

The Q1 is an 88-centimeter, 22-degree-of-freedom developer and companion robot with open control interfaces and customizable shells.
The T1 can switch between wheeled humanoid and quadruped forms, climb stairs, carry objects, and serve as a mobile camera platform.
The company says its line can produce one robot every 2.5 minutes and is using automotive quality processes, a crucial distinction from prototype-heavy competitors.
The launch exposes the real consumer-robot bottleneck: reliable manufacturing, component supply, safety certification, service, and pricing matter as much as locomotion demos.
Q1 still lacks a dexterous hand, showing that a sellable household robot can arrive before general-purpose manipulation is solved.
The strategic test is whether consumers buy useful companionship and developer access, or treat the machines as expensive toys.

Open at Leiphone Ask Claude ChatGPT

Xi lands in Washington for a high-stakes summit

# SCMP

Xi's first US summit visit in 11 years begins with an unusually warm tarmac greeting, even as trade, technology, and security disputes remain unresolved.

Also: ChinaTalk

Trump's personal welcome and the extension of a trade truce provide immediate optics of managed competition rather than reconciliation.
The underlying agenda includes artificial intelligence, critical materials, and strategic rivalry, so ceremonial warmth should not be read as policy convergence.
For Beijing, the visit demonstrates that leader-level diplomacy can preserve channels even when lower-level competition keeps intensifying.
The summit's practical test is whether both sides can convert symbolism into durable guardrails and limited agreements.

Open at SCMP Ask Claude ChatGPT

Alibaba bets on a full-stack economy for machine thought

# GeekPark

Alibaba CEO Eddie Wu framed AI as an industrial system in which chips produce tokens, models determine their quality, and cloud infrastructure distributes machine reasoning.

The strategy treats AI as a production, transport, and consumption chain rather than a single model product.
Alibaba's planned investments in chips, interconnects, models, and global data-center capacity are positioned as control over the supply of machine intelligence.
The economic framing matters because it asks who sets the price, measures output, and owns the distribution network when thought becomes a scalable service.
This is a strategic worldview, not evidence that the company's forecasts about machine reasoning or demand will materialize.

Open at GeekPark Ask Claude ChatGPT

China approves first CAR-T therapy for solid tumors

# SCMP

Chinese regulators approved the world’s first CAR-T therapy for solid tumors, developed by a Shanghai company, and a foreign patient became the first international recipient.

CAR-T has historically been approved mainly for blood cancers, so a solid-tumor authorization would represent a major expansion of the modality’s clinical scope.
The first international treatment occurred at a Shanghai hospital one month after approval, giving China a visible role in both regulatory and medical access.
The report does not provide trial size, response rates, survival data or the target antigen, so the approval should not be read as proof of broad efficacy.
Solid tumors are difficult for CAR-T because of antigen heterogeneity, physical access, immunosuppressive microenvironments and potential off-target toxicity.
The key signal for executives is regulatory and translational: China is willing to authorize advanced cell therapy while building a medical-tourism and international-patient channel around it.
Independent publication of clinical outcomes will determine whether this is a genuine therapeutic breakthrough or an early, narrow approval.

Open at SCMP Ask Claude ChatGPT

US-China investment disputes stall Xi visit business delegation

# SCMP

Disagreements over possible investment announcements have reportedly blocked a Chinese business delegation from joining Xi Jinping’s Washington visit, exposing limits beneath summit pageantry.

The dispute concerns what business commitments can be announced and under what terms, according to sources cited by the report.
A missing business delegation would reduce the summit’s ability to produce visible commercial deliverables even if leaders reach a political understanding.
The episode shows that investment announcements are now part of bilateral bargaining rather than a neutral backdrop to diplomacy.
The report is source-based and incomplete on the specific board or investment proposals, so the direction of the disagreement is clearer than its final outcome.

Open at SCMP Ask Claude ChatGPT

China launches satellites for space-based radio tracking

# SCMP

China placed three Diting-01 satellites into orbit as part of a larger launch, advancing a space-based radio-tracking capability with potential electronic-warfare applications.

The satellites are designed to identify and locate radio emissions from orbit, a capability relevant to surveillance and electromagnetic-spectrum operations.
Their inclusion among nine satellites on a commercial rocket suggests China is building the launch and constellation cadence needed for distributed sensing rather than a one-off experiment.
The report offers no verified performance data, so the significance lies in the mission type and deployment pattern rather than demonstrated battlefield effect.
Space-based radio tracking could complement China’s broader effort to integrate satellites, communications and computing into military and civilian networks.
The system’s real value will depend on revisit rate, geolocation accuracy, signal processing and resilience against deception or jamming.

Open at SCMP Ask Claude ChatGPT

China’s embodied-AI boom confronts a data-center bubble

# TMTPost

A public dispute between robotics executives exposes concerns that government-backed data-collection centers may recycle funds into artificial revenue without end customers.

The reported circular pattern runs from local government investment to a joint data company, then to robot purchases and data buybacks, creating revenue for vendors without proving demand for deployed robots.
Data centers are not inherently fraudulent: physical-agent training genuinely needs large volumes of real interaction, while the industry has far less high-quality data than it claims to need.
The tension is visible in conflicting market signals: the source cites 935 billion yuan (about $132B) of first-half financing and 322 deals, while public-market prices have cooled after spectacular debuts.
This is a governance problem as much as a technical one because local officials receive industrial-development credit, vendors receive orders and investors receive a growth narrative.
The useful diligence question is not how many hours were collected but how much data survives quality filtering, transfers across embodiments and improves a paying customer's task success.
If regulators begin separating related-party revenue, genuine customer orders and reusable data assets, the sector could consolidate quickly.

Open at TMTPost Ask Claude ChatGPT

China's only-child generation begins planning for parental care

# 36Kr

A survey of 878 Chinese only children shows support still flowing from parents to adult children even as more than 80 percent begin supporting aging parents.

Only 20.4 percent of respondents said they received essentially no parental support after adulthood, while more than 80 percent provide some support to parents.
Nearly half handle appointments, medication, or medical reports when parents are ill, and 35.3 percent buy health-monitoring devices.
The financial asymmetry is important: adult children commonly spend small amounts on care, while parents' support is more likely to cover housing, marriage, or other large expenses.
The survey captures the structural burden of China's one-child family system: care, money, and health monitoring are concentrated into a small number of adult children.
For employers and consumer services, this is a demographic shift toward workers managing two generations' needs simultaneously.

Open at 36Kr Ask Claude ChatGPT

AI is industrializing the most labor-intensive step of romance scams

# TMTPost

Chinese reporting links AI chat, fake identities and translation to a lower-cost scam pipeline, while warning that human trafficking and organized crime remain central to the operation.

The report describes AI handling the trust-building conversation while humans still direct payments and the final extraction step, making the scam chain modular rather than fully autonomous.
OpenAI's reported case involved fake profiles, generated images and an AI flirtation agent that moved targets to Telegram; OpenAI said the scale estimates came from the operators and could not be independently verified.
The economics are significant: Chainalysis reportedly found AI-linked scam operations averaged $3.2 million per operation versus $719,000 for others, with 35.1 daily transfers versus 3.89, though attribution is difficult.
Automation may reduce reliance on large scam compounds while increasing the need for live operators at video, payment and coercion stages, potentially making detection harder.
The labor dimension matters: Southeast Asian scam compounds have relied on trafficked and forced labor, and AI may displace or reassign those workers without dismantling the criminal organizations.
The Chinese framing emphasizes industrial process and human exploitation rather than treating the model as an autonomous villain, which places responsibility back on platform operators and criminal networks.
For security and trust teams, the relevant controls are identity verification, payment friction, cross-platform abuse detection and rapid disruption of synthetic relationship infrastructure.

Open at TMTPost Ask Claude ChatGPT

China is assembling a data center in orbit

# TMTPost

China’s satellite projects are combining onboard AI, distributed scheduling and laser links, moving from isolated compute payloads toward an orbital computing network.

The San-Ti computing constellation reportedly launched 12 satellites with up to 744 TOPS per satellite and later linked six satellites under a space-based distributed operating system.
An onboard astronomy model reduced daily downlinked data from hundreds of megabytes to tens of kilobytes by classifying events in orbit, a direct edge-computing use case.
Other projects are testing container updates, incremental model updates and cloud-native management, which are more important for scale than peak TOPS alone.
The Pengcheng-South University satellite combines 5G non-terrestrial networking, onboard compute and laser communications, with a planned 100 Gbps link test.
The system still faces power, thermal, radiation, visibility-window and link-reliability constraints that terrestrial data centers avoid.
China’s strategic direction is clear: treat satellites as schedulable compute and network nodes rather than bespoke spacecraft whose software is frozen before launch.

Open at TMTPost Ask Claude ChatGPT

DeepSeek publishes its large-scale sandbox system for agent training

# IT Home

DeepSeek’s DSec system can create more than 5,000 agent sandboxes per second, exposing the infrastructure needed to train agents across containers and virtual machines.

The reported system reaches about 3 million sandboxes per day and 380,000 concurrent instances on a cluster with roughly 160 nodes, 30,000 CPU cores and 250 TB of memory.
Dsec supports four isolation levels: stateless function calls, containers, Firecracker microVMs and full QEMU virtual machines, all behind one Python interface.
Its environment model separates base images, workspaces and tool packages into independently versioned read-only layers, reducing rebuild cost when one component changes.
Demand loading from DeepSeek’s distributed 3FS filesystem avoids pulling entire images when agents touch only a small fraction of their contents; the paper reports faster burst deployment than cold Docker pulls.
Virtio-pmem and direct access are used to avoid duplicate page caches between microVMs and hosts, cutting peak memory use by 40.2% in the reported test.
The deeper implication is that agent training is becoming a systems problem: isolation, image distribution, scheduling and observability may matter as much as the policy-learning algorithm.
This is unusually valuable because it reveals the operational substrate behind long-horizon coding and computer-use agents rather than presenting only a benchmark score.

Open at IT Home Ask Claude ChatGPT

China reaches 24.22 million EV charging points

# IT Home

China had 24.223 million charging points by August, up 39.6% year on year, with private installations growing more than twice as fast as public ones.

Private points reached 19.063 million and grew 46.3%, while public points reached 5.160 million and grew 19.5%.
Public infrastructure carried 262 million kilowatts of rated power, or about 50.69 kilowatts per point on average.
The gap between private and public growth shows that China’s EV system is being built through household and fleet deployment as much as through visible public stations.
The national plan targets more than 40 million charging points by 2030, enough to support over 110 million electric vehicles.
The hard constraint is shifting from total charger count to grid capacity, geographic availability, peak-load management and high-power charging reliability.

Open at IT Home Ask Claude ChatGPT

Liu Dayiheng takes charge of Alibaba’s Qwen program

# AI QbitAI

Alibaba appears to have appointed early Qwen researcher Liu Dayiheng as its new model lead, consolidating pretraining, post-training and coding responsibilities under a longtime contributor.

Liu reportedly worked on early Qwen pretraining, later led successive model families and helped release more than 300 Qwen variants, making the appointment a continuity choice rather than an outsider hire.
The move follows a reorganization into Alibaba's Token Foundry and months in which Liu expanded from pretraining into post-training and coding teams.
His public theme, real-world agents, points toward a strategy focused on tool use and execution rather than only benchmark model scaling.
Leadership changes near a major Chinese model program matter because they reveal which technical layer Alibaba expects to be its next bottleneck.
The source is a Chinese tech report citing conference materials and does not include a formal corporate appointment announcement.

Open at QbitAI Ask Claude ChatGPT

Jev and Decitron represent two different kinds of decision AI

# AI QbitAI

A Chinese analysis contrasts Jev's fast probabilistic choices for software with Decitron's multi-agent simulation of how decisions change future states.

Jev emits a choice, score or probability that software can use directly, effectively replacing some hand-written routing and threshold logic with calibrated model judgments.
Decitron models state, action and outcome across multiple actors, then compares possible futures through world modeling, games and optimization.
The distinction is architecturally useful: one system is a low-latency decision primitive, while the other is a planning and simulation environment.
Both make uncertainty explicit, but one measures confidence in the next choice while the other represents uncertainty over how the world may evolve.
The Chinese industry signal is a move beyond text generation toward machine-readable decisions and action policies, though the source offers no benchmark or deployment evidence.

Open at QbitAI Ask Claude ChatGPT

DeepSeek reveals how it runs agent reinforcement learning

# AI QbitAI

DeepSeek's DSec paper explains how it creates thousands of isolated training environments per second for code, security, browser, and full-computer agents.

The platform unifies function calls, containers, microVMs, and full virtual machines behind one interface because agent tasks require different isolation and operating-system fidelity.
Layered read-only images and demand loading reduce deployment time and storage writes when thousands of environments launch at once.
Memory mapping and cold-page reclamation let many sandboxes share environment data instead of duplicating it in host and guest caches.
The paper makes clear that the bottleneck in agent RL is often the environment lifecycle, not just model inference or GPU throughput.

Open at QbitAI Ask Claude ChatGPT

Guangzhou stabbing case fuels anger over censorship and official silence

# Culture China Digital Times

After a man appeared to target women outside a Guangzhou shopping center, Chinese online discussion criticized deleted posts, limited official detail and the routine reliance on brief police notices.

The source says police issued a short statement while users and bloggers noted that discussion and images were removed or suppressed online.
A deleted essay argued that describing the attack as indiscriminate obscured the apparent targeting of women and their faces, turning the incident into a gendered public-safety debate.
The wider pattern described is institutional: authorities provide minimal facts, news outlets wait for official statements and independent follow-up becomes difficult.
This is a window into how censorship and information control shape public understanding of violent incidents, not a confirmed account of motive or the full victim count.
The episode matters because online users are contesting not only the attack but also the official vocabulary used to describe it.

Open at China Digital Times Ask Claude ChatGPT

China says purging two military chiefs removed a political danger

# Policy SCMP

The military's official newspaper says the expulsion of former Central Military Commission figures Zhang Youxia and Liu Zhenli eliminated a major hidden political danger.

The two former senior officers were expelled from the Communist Party and military over alleged factionalism and disloyalty, according to state media cited by the source.
Because the Central Military Commission is chaired by Xi, the purge touches the top of the command structure rather than a routine personnel case.
The official phrase about removing a political danger signals that the leadership is framing the move as a threat to Party control and military loyalty, not merely corruption.
The timing alongside a broader reshuffle suggests continued consolidation of authority and possible concern about factional networks inside the People's Liberation Army.
The source does not establish the underlying evidence or operational consequences, so the official explanation should be read as political messaging as well as fact reporting.
For technology and defense observers, the key question is whether the purge changes procurement, command continuity or the military's relationship with the civilian industrial system.

Open at SCMP Ask Claude ChatGPT

China policy watchers parse Xi’s summit agenda and military expulsions

# Policy Sinocism

Sinocism links Xi’s Washington agenda with the September Politburo’s formal expulsion of generals Zhang Youxia and Liu Zhenli, giving the summit an internal-security backdrop.

Also: SCMP

The juxtaposition matters because external diplomacy and military personnel politics are being watched together: a confident summit posture can coexist with internal discipline and uncertainty.
The newsletter also flags unresolved questions about an AI hotline, a trade mechanism and rare-earth friction, suggesting the bilateral relationship is being managed through partial channels rather than a grand settlement.
Claims about the generals and their meaning should be treated as newsletter analysis unless supported by formal Party documents.
The source is valuable for showing what China watchers are prioritizing, but it is not a primary account of Politburo deliberations.

Open at Sinocism Ask Claude ChatGPT

Voyah launches a 420,000 yuan flagship MPV with Huawei driving tech

# AI QbitAI

Voyah’s Dreamer 9 starts at 419,900 yuan, roughly $59,000, combining Huawei’s ADS 5, four lidar units, an 800-volt system and a redundant L3-ready architecture.

The vehicle uses a 896-line main lidar, three solid-state blind-spot lidars and 37 sensing components for urban, highway and parking scenarios.
Its chassis combines three-chamber air springs, dual-valve dampers, vehicle-motion control and 10-degree rear-wheel steering to manage a 5.3-meter, high-mass MPV.
The architecture includes redundant sensing, computing and power systems, but the article correctly notes that actual L3 availability still depends on regulation and delivered software.
A 560-kilowatt plug-in hybrid version offers a claimed 340-kilometer electric range and 12-minute 20% to 80% charging, showing how Chinese luxury vehicles are bundling software, chassis and powertrain differentiation.
The strategy is to compete with imported luxury MPVs by making system integration, not just cabin space, the flagship feature.

Open at QbitAI Ask Claude ChatGPT

PixVerse R2 claims real-time world-model scaling

# AI QbitAI

Chinese startup AIsphere says PixVerse R2 combines persistent spatial state, text and audio control, and real-time generation in one world-model framework.

The technical claim is a unified causal representation for visual, action, audio and temporal signals, intended to let scenes evolve consistently rather than regenerate disconnected frames.
The model aims to preserve state and respond to prompts while generating, addressing the usual tradeoff between richer world understanding and real-time latency.
The source is promotional and supplies no latency, hardware, benchmark or failure-rate data, so the claimed generality remains unverified.
If the architecture scales beyond demonstrations, interactive games and embodied simulation could become new consumers of generative video infrastructure.

Open at QbitAI Ask Claude ChatGPT

Agents are starting to take technical decisions from developers

# Analysis InfoQ China

A Chinese essay argues that coding agents are moving from code completion to choosing languages, frameworks and workflows, shifting technical power away from developers.

The argument is not that human developers disappear immediately, but that software decisions increasingly emerge from model defaults, prompts, tool interfaces and agent feedback loops.
The economic analogy is that token cost may replace part of labor cost, while agent-readable interfaces such as command lines, skills and tool protocols become strategic distribution channels.
This raises governance questions for engineering leaders: who approves dependency choices, how are prohibited technologies blocked, and how can decisions be audited after autonomous execution?
The essay's strongest claim is still speculative, but the organizational direction is visible in products that let agents write, test, revise and ship with limited human intervention.

Open at InfoQ China Ask Claude ChatGPT

Lexiang’s embodied model tests robots in an uncontrolled barbecue stall

# AI Leiphone

Lexiang used a live, interruptible barbecue task to show robots handling orders, memory, physical manipulation and recovery without a scripted demonstration.

The robots had to coordinate ordering, food preparation, serving and cleanup while viewers changed layouts, objects and requirements, making the test closer to an open-loop service environment than a staged demo.
One robot reportedly handled an interruption, delivered water and resumed the original order; another adjusted its grasp after dropping food, suggesting state retention and feedback-driven correction.
The architecture is presented as solving actions from current physical constraints rather than retrieving a memorized motion, which is the key distinction between a demo policy and a general controller.
The event was organized by the company after public controversy over its model and demo claims, so it is evidence of a more adversarial test format but not independent validation.
The useful benchmark question is whether success persists across unseen kitchens, tools, temperatures and human interruptions, not whether one live stream ran smoothly.

Open at Leiphone Ask Claude ChatGPT

Xiaomi's MiMo V2.6 trains agents across many environments

# AI Leiphone

Xiaomi's open MiMo V2.6 combines sparse attention, mixture-of-experts routing, speculative decoding, and asynchronous reinforcement learning across coding, vision, agents, and security.

The Pro model has 1.02 trillion total parameters with about 42 billion active per token; most layers use 128-token sliding windows while a minority provide global attention.
One RL update reportedly samples 1,568 prompts with 16 rollouts each, creating 25,088 trajectories whose different runtimes make synchronous training wasteful.
Fully asynchronous GRPO separates rollout, environment execution, grading, and parameter updates so workers can continue rather than wait for the slowest trajectory.
Training several task families and harnesses together aims to transfer state understanding, tool choice, and recovery behavior instead of overfitting to one interface.
The architecture is a concrete example of agent training becoming a systems problem involving schedulers, environments, and reward design as much as model weights.

Open at Leiphone Ask Claude ChatGPT

Alibaba wants to be the operating layer for Chinese AI phones

# AI Zhidx

Alibaba's Qwen Intelligence combines models, agents, execution harnesses, cloud-device coordination, and security for phone makers rather than building a handset itself.

The stack offers planner, mobile-use, and creative agents, with an API-first and GUI-fallback strategy for apps that lack agent interfaces.
Alibaba is targeting the gap between a model demo and a completed task: state, memory, permissions, tool calls, and feedback must be coordinated around the operating system.
The commercial strategy is platform infrastructure for phone brands, with Honor as an early integration partner rather than a consumer-facing Alibaba device.
The challenge is ecosystem cooperation: an agent can reason correctly yet fail if apps expose no stable APIs or if privacy controls block execution.

Open at Zhidx Ask Claude ChatGPT

Huawei veterans launch a physics foundation-model startup

# AI Zhidx

Chinese startup Xirra raised several hundred million yuan (roughly $30-60 million) in seed and angel funding to build a physics foundation model for robots and autonomous systems.

The founders are a former Huawei cloud-model executive and a causality researcher who are combining large-scale training experience with physical-world modeling.
Unlike the common approach of adding an action head to a vision-language model, the company is attempting native pretraining on state, action, and outcome relationships.
Its proposed architecture separates causal world prediction from video rendering, using autoregressive reasoning for what happens and diffusion for how it looks.
The technical challenge is severe: physical state is continuous and partially observed, trajectories are long, data are heterogeneous, and simulation differs from reality.
The funding shows Chinese investors are willing to back a foundational layer for embodied AI, but the supplied performance evidence is limited to company and benchmark claims.

Open at Zhidx Ask Claude ChatGPT

China tightens fentanyl rules ahead of the Xi-Trump summit

# Policy SCMP

Beijing announces stricter fentanyl controls as Chinese officials describe Xi's US visit as a historic milestone, combining a policy gesture with summit diplomacy.

The timing makes the regulation part of diplomatic signaling as well as domestic drug control, giving Washington a concrete action to cite before the leaders meet.
Chinese official language frames the relationship as a large vessel that should keep moving forward under leader-level guidance, emphasizing stability and mutual management.
The strategic value depends on implementation and measurable cooperation, not the announcement alone; precursor controls and enforcement across supply chains matter more than rhetoric.
The episode shows how Beijing often pairs a policy adjustment with carefully staged political messaging when seeking momentum in a high-level relationship.

Open at SCMP Ask Claude ChatGPT

Washington weighs Beijing’s bigger trade proposal before Xi visit

# Policy SCMP

US Treasury Secretary Scott Bessent says Washington may accept a broader Chinese trade deal or extend last year's Busan agreement before Xi's arrival.

The negotiation shows both sides using a temporary trade framework as a floor while testing whether a larger bargain is politically and economically possible.
Chinese leverage appears strongest where supply-chain dependencies and rare-earth concerns meet US tariff pressure, but the source excerpt does not specify the proposed terms.
The meeting is substantive China-US bargaining rather than ceremony, making implementation and enforcement more important than the label attached to the deal.

Open at SCMP Ask Claude ChatGPT

A celebrity scandal exposes China's gendered attention economy

# Culture Initium

A viral post by crypto entrepreneur Justin Sun about actress Jing Tian triggered backlash, showing how wealth, intimacy and misogyny are negotiated in Chinese online culture.

The source says Sun's unverified, explicitly fictionalized essay used extravagant spending and a dispute over 300 million yuan, roughly $42 million, to make crypto wealth legible to a mass Chinese audience.
The controversy was not only about the claims' truth; it turned a private relationship into a public performance of male purchasing power and sexual status.
Young women broadly supported Jing Tian because intimate details do not produce equal reputational costs: men can present a partner as an achievement while women face voyeurism and slut-shaming.
The episode shows how China's online public sphere can combine celebrity gossip, crypto wealth, class resentment and feminist counter-speech in one rapidly moving scandal.
Because the central account was not substantiated, its importance is cultural and political rather than evidentiary.

Open at Initium Ask Claude ChatGPT

Damao Technology raises about $280 million for compute-power coordination

# Policy QbitAI

Chinese energy-AI company Damao Technology reportedly raised about 2 billion yuan (roughly $280 million) for software coordinating data centers with renewable power and electricity markets.

Its platform treats data centers as flexible electrical loads that can shift tasks according to renewable output, prices, and service constraints.
The company combines energy models, virtual power-plant software, and long-term operations rather than selling a standalone dashboard.
The story reflects a policy direction in which compute-power coordination and direct renewable connections are being elevated as national infrastructure priorities.

Open at QbitAI Ask Claude ChatGPT

Beijing brought a fentanyl rule and a Musk testimonial to Washington, and left its own CEOs at home

The welcome is bigger than the agenda

The chip minister gets the chip province

Frontier AI has become an infrastructure bill

AI took over the expensive step in a pig-butchering scam

China is now where elite AI researchers go

Threads we are pulling