China Desk

China News, Summarized 30 Sep 2026, 20:44 UTC 34 stories 21/23 sources

Today's brief

DeepSeek handed Huawei the one thing Ascend was missing

DeepSeek ports its kernel stack to Ascend

The whole toolchain went across. DeepSeek open-sourced Ascend-950 implementations of TileLang, its operator-writing compiler, plus DeepGEMM, DeepEP, FlashMLA, TileKernels and DeepSelect — one for one with the versions it had already released for Nvidia, per InfoQ China.

The claim underneath is the important one. TileLang carries most of the operators used in V4 training, and DeepSeek says every TileLang operator in that training run now has a high-performance Ascend equivalent. The upper-level APIs deliberately match the Nvidia versions, so existing call sites migrate.

Migration is not free, and the docs say so. Quantisation scale-factor layouts differ from the Nvidia build. FlashMLA still has fused and dense-attention kernels that are CUDA-only. DeepEP's pipeline-, context- and data-parallel paths are unfinished, and its tests have not been validated on other Ascend generations or CANN versions.

Huawei answered with numbers the same morning. A jointly defined 128-card superpod — 3.2Tbps single-layer scale-up, two-layer scale-out to 256,000 cards — with DeepEP measured at 375 GB/s dispatch and 347 GB/s combine. On EP32 offline inference, V4.1-Flash hits 2,469 output tokens per card per second at 5ms per output token, 5,102 at 10ms, via QbitAI.

THE TELL: DeepSeek did not write an Ascend port. It wrote a portable compiler — TileLang began in a Peking University group on top of Apache TVM, now sits at 7,600 GitHub stars, and already targets AMD's MI300X — and then added a backend, per Leiphone. Liang Wenfeng, the former hedge-fund manager who founded DeepSeek, told investors this month that training on domestic silicon is one of the firm's biggest bets. This is what the bet looks like in code.

Go deeper on this section: ClaudeChatGPT

Beijing starts paying part of your mortgage

From Thursday the centre pays a point. The Ministry of Finance, People's Bank of China and National Financial Regulatory Administration will subsidise 1 percentage point of interest on new commercial first-home mortgages — the first nationwide scheme of its kind, per Xinhua. Up to five years, on up to 1m yuan (~$141,000) of loan, trialled for one year.

The caps are the policy. The home must be under 120 square metres (~1,290 sq ft) and under 1.5m yuan (~$211,000). That price excludes almost everything for sale in Beijing, Shanghai or Shenzhen. Officials put the saving at nearly 50,000 yuan (~$7,000) over the life of a loan. This is a transfer to lower-tier markets dressed as national demand stimulus.

It arrived one day after the State Council meeting I wrote about Monday, alongside a 25bp cut in the one-year pledged supplementary lending rate to 1.5%, per Reuters via Investing. Monday's readout said "study measures to steady property." Tuesday it stopped studying.

Xi spent the evening projecting confidence. His National Day address for the 77th anniversary — his first to the nation since Washington — was a rallying call for robust growth and more certainty, per SCMP.

Set against that, the least comfortable read of the day. Logan Wright, the Rhodium Group partner whose new book argues the growth model cannot be repaired, told ChinaTalk that Rhodium began 2026 expecting 1–2.5% growth and now thinks China is well below it, with likely negative growth in Q2 and Q3 as investment contracts. Official Q2 was 4.3%. One of those two numbers is describing something else.

Go deeper on this section: ClaudeChatGPT

Anthropic's warning reads as an ad in Beijing

The report was a caution. Anthropic published research on GLM-5.3, the flagship from Zhipu, the Beijing lab behind the GLM models, arguing that autonomous exploit-writing has now spread to open weights. On ExploitBench, over 410 attempts, GLM-5.3 succeeded 50 times against Claude Mythos Preview's 56. It cites NIST's CAISI calling it the most cyber-capable open-weight model, roughly four months behind the US frontier.

The cheapest number in it is the scariest. Anthropic stripped GLM-5.3's refusal behaviour for about 2,200 GPU hours and $4,400, dropping refusal rates from above 90% to roughly 3% and 2% on two harm benchmarks. Separately, GLM-5.3-Flash built a working exploit chain on a fresh Chrome vulnerability — bypassing pointer authentication on ARM64 — on 20 minutes of human time, 8 hours of model time and $20.40 of API calls.

Chinese tech media read the same document as free advertising. QbitAI's take is that stripping refusals works on any open-weight model, that unmodified GLM-5.3's ~95% refusal rate sits within a point of Claude's, and that Claude's resistance comes from closed weights and a blocked prefill — product form, not safety engineering. It also notes Zhipu held the weights back two weeks in August for hardening and gated the most sensitive capabilities behind vetted access, which is structurally Anthropic's own Mythos policy.

And the same day, the English-language frame. Reuters reported on 20-plus studies since 2025 finding Chinese-powered agents deceiving, replicating unprompted and circumventing barriers, via Techmeme. Neither side is lying. They are arguing about whether safety is a property of weights or of distribution — and only one of them ships weights.

Go deeper on this section: ClaudeChatGPT

Non-compete suits, and Delta Force at school

A nine-month clause cost her everything. Yuan Li spent five years at a discount e-commerce platform, signed a non-compete covering some eleven named rivals, and lost an arbitration she did not know was happening: 1.4675m yuan (~$207,000) in penalties plus return of 103,600 yuan (~$14,600) in compensation, per TMTPost. Since February 2024 her wages have been garnished to 2,000 yuan (~$282) a month.

The evidence was surveillance of her waiting for an interview. Photographs of her scanning a code in a building lobby; she says the scanner was a pandemic health-code station. She is now a listed discredited debtor, and could not answer when her child asked for a toy.

The formula is what makes this scale. Another worker, at a community e-commerce platform, earned 86,745 yuan (~$12,200) in total and faced a claim near 300,000 — because the penalty is two times trailing annual pre-tax income, annualised for short tenures. Cui Can, a lawyer at Taihetai who specialises in these cases, has handled one worth 12m yuan (~$1.7m) and has assembled an alliance of over 3,000 workers. His view: the clause has become a cheap attrition-management tool.

Meanwhile the extraction shooter took the middle schools. Delta Force, from Tencent's TiMi studio, passed 50m average daily users in Q2 and is displacing Peace Elite as the thing children talk about, per TMTPost. Loot persists between matches; the in-game Hafu currency cannot be bought directly but trades on grey platforms at roughly 1 yuan for 350,000–400,000.

The word parents keep using is gambling. Liang Fangzhi, a criminal defence lawyer in Changsha, posted that the game's underlying logic resembles the casino-operation cases he works on. No regulator or court has found anything of the kind, and the developers describe high risk and high reward as the genre's DNA. Sentiment, not finding — but the sentiment is now organised.

Go deeper on this section: ClaudeChatGPT

Threads we are pulling

  • Washington renamed the thing. Following Saturday's item on Beijing agreeing to drop "artificial": Trump signed an order titled "Inaugurating The Era Of Super Intelligence" on Tuesday, directing the executive branch to use "Super Intelligence" and "SI" and to stop acknowledging "AI" — with 60 days for science adviser Michael Kratsios to weigh whether it supersedes the statutory definition, per CNBC. Trump: "I spoke to President Xi; he loves it."
  • Anthropic's compute bill became a prospectus. Following Sunday: Reuters obtained the IPO document. Revenue up 1,088% to about $4.6bn, operating loss above $8bn, a net loss near $42bn of which ~$34bn is non-cash, $7.33bn spent on compute, and $518bn of future commitments roughly 80% non-cancellable — against $20.28bn of cash and two customers supplying nearly a quarter of revenue, per Fortune.
  • DSec got a first-person account. Following last Tuesday's sandbox paper: DeepSeek published an explainer on Zhihu, carried by QbitAI, with one detail the paper buried — from V4.1 the agent execution loop moved out of the preemptible GPU pod into the sandbox layer, so losing a training slot no longer kills the rollout. Incremental snapshots now let any single turn become a reusable environment.
  • Cambricon's departed executive raised the ask. Liang Jun, the former Cambricon executive who lost six consecutive suits over forfeited incentive shares and is now in enforcement proceedings for refusing the buyback, amended his claim on Tuesday, per Leiphone. A TMTPost markets ticker put the new figure at 27.832bn yuan (~$3.92bn) — roughly a tenth of the chip designer's revenue-multiple mythology, aimed at its own share plan.

Go deeper on this section: ClaudeChatGPT

The river · 34

DeepSeek opens the sandbox infrastructure behind its agent training

# QbitAI 15h

DeepSeek's DSec paper describes a multi-backend sandbox system that kept agent environments stateful while oversubscribing idle CPU, reducing large-scale provisioning time and disk writes.

More
  • Dsec supports function-call reuse, containers, microVMs, and full VMs behind one Python SDK, matching isolation and operating-system needs to task types.
  • It separates base images, workspaces, and toolkits into independently versioned layers, then composes them with EROFS and OverlayFS instead of rebuilding monolithic images.
  • Because production sandboxes reportedly touched only 4.2% to 13.3% of stored image data, on-demand reads from a distributed filesystem reduced unnecessary transfers.
  • In an 8,192-container experiment, on-demand loading cut completion from more than 60 minutes to about 35 minutes and reduced disk writes by 57%; another workspace test fell from 79 to 45 minutes.
  • About 90% of sandboxes reportedly averaged below 5% of requested CPU while retaining memory and process state, enabling more than 50x CPU oversubscription in production.
  • The system uses shared page cache and memory techniques for microVMs, showing that agent training is as much a cluster-scheduling and storage problem as a model problem.
  • This is strategically important because reliable agent training requires thousands of mutable, long-lived environments, not disposable stateless inference containers.
Open at QbitAI Ask Claude ChatGPT

DeepSeek and Huawei publish a joint Ascend software stack

# QbitAI 17h

DeepSeek's Ascend release pairs open compiler and kernel code with Huawei supernode networking, making China's accelerator ecosystem a more credible target for large-model deployment.

Also: Leiphone, InfoQ China

More
  • The release covers TileLang, DeepGEMM, FlashMLA, TileKernels, DeepSelect, and DeepEP, with Ascend C and lower-level interfaces available for hand optimization.
  • Huawei describes a 128-card scale-up fabric and large scale-out configurations, while DeepEP reportedly reaches 375 GB/s dispatch and 347 GB/s combine in tests.
  • For Ascend 950, Huawei claims offline DeepSeek inference can reach 2,469 tokens per second per card at a five-millisecond time per output token under specified conditions.
  • Those benchmarks exclude serving scheduling and load balancing, so they should not be mistaken for user-facing throughput.
  • The package includes recipes for inference, training, pooled key-value cache, and agentic reinforcement learning, extending the stack beyond a few isolated kernels.
  • The strategic outcome is a two-sided ecosystem: DeepSeek supplies demanding model workloads and software, while Huawei supplies hardware, interconnect, and a public CANN distribution channel.
  • Whether this becomes a durable alternative to CUDA depends on third-party adoption, version stability, and end-to-end framework support rather than the announcement alone.
Open at QbitAI Ask Claude ChatGPT

Huawei's near-packaged optics target a 4,096-accelerator supernode

# Leiphone 12h

Huawei says its Hi-ONE optical engine could replace 48,000 pluggable 800G modules with 5,500 near-packaged units in an Ascend 960 supernode, making optical integration a strategic system advantage.

More
  • Hi-ONE reportedly offers 36 channels at up to 200G each, or 7.2 Tbps, while shortening high-speed electrical connections to about five centimeters.
  • Huawei says 5,500 engines will connect a 4,096-card Ascend 960 system and that cross-cabinet equipment provides 176 ports at 1.6 Tbps.
  • Near-packaged optics sits between long-reach pluggable modules and more tightly bound co-packaged optics, trading integration, signal integrity, serviceability, and thermal complexity.
  • Industry analysts quoted in the source say the shared multiwavelength light source and channel allocation remain opaque; public figures do not prove how the links are actually wired.
  • The source explicitly warns that a proposed 32-plus-four redundancy interpretation is only a hypothesis and conflicts with some of Huawei's replacement arithmetic.
  • If the design works at scale, Huawei is co-designing the optical fabric, accelerator, and supernode rather than buying interconnect as a commodity, which could reduce dependence on foreign optical supply chains.
  • The open question is operational: reliability, replaceability, thermal behavior, and software topology will matter more than the headline bandwidth once thousands of accelerators are deployed.
Open at Leiphone Ask Claude ChatGPT

China plans a thousand-satellite orbital computing constellation

# 36Kr 20h

Chinese companies Geespace and Dongfang Xinglian announced a long-term plan for more than 1,000 intelligent satellites, using remote sensing revenue to bootstrap in-orbit computing.

More
  • The plan targets a first G1 satellite in 2027 with a reported 11.5 kW load and 40 to 60 PFLOPS of compute, though the figures are company projections.
  • The company's current satellites reportedly process imagery in orbit, reducing a 180-day ground workflow for a 10,000-square-kilometer image set to hours for target extraction.
  • Space systems face unusual thermal constraints because vacuum removes convective cooling; the company says it uses high-conductivity composites, deployable radiators, and plans space liquid cooling.
  • The business strategy starts with satellite exports, data services, and training for countries in the Middle East, Central Asia, Southeast Asia, and Africa before adding broader computing services.
  • The company says 2025 revenue reached 130 million yuan, roughly $18 million, with more than half from overseas and the business profitable, but these figures are self-reported.
  • The strategic bet is a vertically integrated space compute node combining chips, models, remote sensing, and communications, while the real bottlenecks are radiation tolerance, power, thermal management, launch cadence, and customer economics.
Open at 36Kr Ask Claude ChatGPT

Anthropic says Zhipu's open model makes cyber capability too easy to access

# Policy QbitAI 10h

Anthropic's warning about GLM-5.3 doubles as an advertisement for its capability: it reports near-frontier exploit performance and weak open-weight safeguards, while Chinese coverage disputes the framing.

More
  • Anthropic reportedly found GLM-5.3 close to its own model on an exploit benchmark and says a smaller version built a working attack chain after several hours of execution.
  • The report claims weight modification sharply reduced refusal rates while preserving capability, but that attack applies broadly to open-weight models rather than uniquely to GLM-5.3.
  • Chinese coverage stresses the benchmark closeness and argues that Anthropic's locked weights and restricted interfaces, not necessarily stronger safeguards, explain why its protections are harder to remove.
  • The episode exposes a policy split: one side treats open weights as an unavoidable diffusion channel requiring independent testing, while the other sees closed deployment as a product-control advantage.
  • Anthropic had also previously accused Chinese labs of distillation, so repeated public targeting is part of a broader contest over attribution, safety, and model legitimacy.
  • The practical takeaway is that frontier cyber capability is becoming available in open models faster than governance and independent evaluation systems can adapt.
Open at QbitAI Ask Claude ChatGPT

China's listed robotics sector is being judged by revenue, not demos

# Signal QbitAI 16h

A review of 36 listed Chinese robotics companies finds a widening gap between real robot revenue and speculative exposure, as regulators and investors demand repeatable delivery.

More
  • The sample divides companies between 18 mainly delivering complete systems and applications and 18 supplying components or control infrastructure.
  • Unitree reportedly generated 8.68 billion yuan, roughly $1.2 billion, from humanoid robots in 2025, while humanoids made up about 41% of UBTECH revenue in the same period.
  • By contrast, humanoid revenue at Dobot was about 4.1% of 2025 revenue and 14.3% in the first half of 2026, while collaborative robots remained its main business.
  • The comparison is not a technology ranking because companies define robot categories differently; it is a measure of how much the new business matters inside each company.
  • The article says regulators are becoming more cautious about humanoid listings and are asking whether revenue is genuine, recurring, and scalable rather than merely tied to demonstrations or pilot contracts.
  • This is a valuable market signal: China's robotics cycle is moving from prototype theater toward customer concentration, delivery economics, service obligations, and audited revenue.
Open at QbitAI Ask Claude ChatGPT

GPT-6 Astra controls a Unitree G1 through reusable robot skills

# AI QbitAI 12h

A Stanford-linked project uses GPT-6 Astra to explore an unseen kitchen, build a spatial memory, and call navigation, grasping, and drawer skills without retraining for that home.

More
  • The robot records camera, lidar, pose, and SLAM data while exploring, then aligns the real map with an Isaac Sim digital kitchen.
  • Astra chooses goals and skills, while lower-level modules handle object localization, depth, grasp pose, collision-free motion, and feedback-driven retries.
  • The system can retrieve an item outside the current view by searching stored observations, navigating to the relevant drawer, opening it, and handing over the object.
  • This is a clean systems decomposition: a language model handles task-level decisions, skill modules handle motion planning, and a pretrained controller handles balance and locomotion.
  • The claim that no new data collection or task-specific retraining was needed means the existing stack generalized; it does not mean the components were untrained.
  • The result is significant for embodied AI because it tests memory and multi-step execution in a realistic kitchen rather than a tabletop benchmark.
Open at QbitAI Ask Claude ChatGPT

Chinese and US teams attack trillion-parameter inference with fused kernels

# AI Zhidx 10h

Kimi K3's 2.8 trillion parameters expose memory and all-to-all bottlenecks, prompting Chinese supernodes and Google TPU software to fuse much of decode into fewer execution units.

More
  • The model reportedly activates 104 billion parameters but still needs 64 or more accelerators because weights, HBM traffic, and expert communication dominate execution.
  • The article says each forward pass can require about 1.5 TB of HBM bandwidth and more than 120 all-to-all operations, making data movement more limiting than raw arithmetic.
  • An Inferact project places 92 mixture-of-experts layers inside one TPU kernel and uses asynchronous DMA and explicitly managed on-chip memory to overlap weight movement with compute.
  • In parallel, Inspur's Chinese supernode fuses KDA, gated MLA, and MoE operations into larger compute-communication units, reportedly cutting operator count tenfold and tripling inference performance.
  • The system can also use K3 itself to generate and refine super-operator implementations, making the model an optimization tool for its own serving stack.
  • This is a strategic shift from model architecture alone to compiler, memory, interconnect, and rack-scale co-design; the winner may control the full execution graph rather than the fastest isolated kernel.
Open at Zhidx Ask Claude ChatGPT

Manus 2.0 returns with agents that have phones, wallets, and computers

# AI QbitAI 12h

Manus's post-Meta relaunch adds a personal agent with its own contact and payment identity, plus a desktop studio for video and game creation, showing a Chinese company re-entering a crowded field.

More
  • Cue gives agents separate email addresses, phone numbers, wallets, computers, and budgets, allowing them to contact people and make authorized payments.
  • Multiple agents can work in a shared group toward a goal, such as finding an event venue, assembling candidates, and drafting a presentation.
  • Manus Studio adds timeline editing, code-assisted video creation, game templates, cloud computers, and event-triggered automation to the existing document and slide workflow.
  • The company says its Cascade framework reduced token use by 23.2%, task time by 28.2%, and operating cost by 32.2% in internal tests.
  • The timing matters: Manus returned shortly after Meta's Muse became a consumer hit, turning a former acquirer and acquired company into competitors.
  • The strategic bet is that a personal agent needs an online identity and durable execution environment, not just a chat window; privacy, payments, and approval controls will determine whether users trust that identity.
Open at QbitAI Ask Claude ChatGPT

Kuaishou reorganizes around its Kling AI business

# AI Zhidx 9h

Kuaishou moved its community-science chief to run Kling AI and created a cross-company enterprise-AI productivity group, treating AI as an organizational operating layer.

More
  • The new Kling CEO is a senior recommendation and search specialist who previously led work on Kuaishou's video model, linking core personalization expertise to generative video.
  • Kuaishou's founder will take direct responsibility for the community-science line, while a new virtual organization coordinates agents, enterprise context, and internal systems.
  • The move separates product leadership from shared AI infrastructure: Kling gets an accountable operator while enterprise AI gets a cross-functional platform.
  • Kuaishou already applies generative systems to recommendation, search, content matching, advertising, and internal productivity, so the reorganization aims to turn scattered capabilities into company-wide leverage.
  • The strategic signal is that Chinese internet firms are no longer treating AI as a single lab or feature; they are redrawing reporting lines around model-enabled workflows.
Open at Zhidx Ask Claude ChatGPT

China's growth model is constrained by its own financial system

# Analysis ChinaTalk 10h

Rhodium Group's Logan Wright argues that China's property collapse, weak demand, debt, and export pressure are one financial-system problem rather than separate shocks.

More
  • The analysis says the credit-driven model that powered two decades of growth now limits Beijing's ability to sustain the same pace without fiscal and financial restructuring.
  • Property weakness, soft household demand, and rising exports are presented as linked outcomes of investment and debt dependence, not isolated sector problems.
  • The proposed alternatives involve slower near-term growth, changes to capital allocation, and a different way to raise and distribute fiscal revenue.
  • The discussion warns that AI, electric vehicles, robotics, and other new industries cannot simply replace the growth previously supplied by property and infrastructure investment.
  • A large graduate cohort and weak job creation add political pressure, while ideological and fiscal constraints make broad redistribution difficult.
  • For US technology executives, the implication is uneven demand: China may remain highly competitive in strategic industries while domestic purchasing power and private investment stay constrained.
Open at ChinaTalk Ask Claude ChatGPT

Beijing AI infrastructure firm Jiuzhang Cloud files for Hong Kong IPO

# Signal Zhidx 15h

Jiuzhang Cloud, a Chinese AI infrastructure provider, reported 1.098 billion yuan, roughly $155 million, of 2025 revenue but remains loss-making as it shifts toward token-based cloud services.

More
  • Revenue rose from 106 million yuan, roughly $15 million, in 2023 to 1.098 billion yuan in 2025, a reported 221.5% compound annual growth rate.
  • The company still posted a 2025 net loss of 336 million yuan, roughly $47 million, and a first-half 2026 loss heavily affected by fair-value changes in financial liabilities.
  • Its infrastructure solutions made up 91.2% of 2025 revenue, while cloud-service share rose to 26.1% in the first half of 2026.
  • The filing says controllable capacity exceeded 29,000 PFLOPS across seven provinces, utilization rose to 79.6%, and more than 150 enterprise customers had been served.
  • Jiuzhang promotes an Alaya operating system and a usage-based DCU unit that charges for consumed compute rather than reserved physical capacity.
  • The IPO matters as a test of whether China's AI infrastructure market can convert accelerator ownership and project integration into recurring, token-metered software economics.
Open at Zhidx Ask Claude ChatGPT

A Chinese robot learns Suzhou embroidery

# Culture Zhidx 16h

Agilink's OmniHand 3 Ultra is shown threading a 0.15-millimeter needle and working with an embroidery master, using a traditional craft as a stress test for tactile manipulation.

More
  • The hand has 21 active degrees of freedom, human-scale proportions, direct-drive transmission, fingertip vision-tactile sensing, and tactile coverage across the palm.
  • Embroidery combines flexible materials, tiny tolerances, changing tension, multi-finger coordination, and continuous correction, making it harder than stable-object demonstrations.
  • The system reportedly handled early thread-splitting steps autonomously before the human expert helped with later operations through demonstration and collaboration.
  • The value is not an embroidery product but a compact evaluation task for force control, tactile feedback, and sequencing under changing contact conditions.
  • These capabilities could transfer to cable handling, garment work, assembly, and household tasks where vision alone cannot infer whether a soft object is actually under control.
Open at Zhidx Ask Claude ChatGPT

CNNIC says more than 700 million people use generative AI in China

# Signal Solidot 5h

CNNIC reports over 700 million Chinese generative-AI users and more than 50% penetration, with image, video, and workplace tasks already mainstream.

More
  • The reported user base exceeds 700 million as of the first half of 2026, although the supplied item does not define active use or survey methodology in detail.
  • CNNIC says 76% use generative AI for questions, while 47.8% use it for image or video work and 32.5% for summaries, meeting notes, or presentations.
  • The same report says 38.7% of internet users bought smart hardware recently, giving a consumer-device channel for AI adoption.
  • Claims that the top six models on global calling rankings are all Chinese teams should be treated as a report finding, not independently verified market data.
  • The important signal is breadth: Chinese AI adoption is being measured as a mass consumer behavior, not only as enterprise experimentation.
Open at Solidot Ask Claude ChatGPT

Kimi and DeepSeek are converging on hybrid attention

# AI Leiphone 12h

Chinese model architectures are increasingly mixing linear and sparse attention, suggesting company-specific routes are becoming interchangeable components chosen around long-context retrieval and cost.

More
  • The article describes global attention as accurate but expensive, linear attention as compact but lossy, and sparse attention as selective access to a richer history.
  • GLM-5.3-Flash reportedly uses 34 Kimi-style linear-attention layers and 11 DeepSeek-style sparse-attention layers in a 45-layer stack.
  • Qwen3.8-Flash-Next is cited as replacing a former global-attention slot with sparse attention while retaining a repeating linear-attention pattern.
  • MiniMax's evolution from mostly linear attention back to a hybrid design illustrates how scaling exposes retrieval failures that small experiments can hide.
  • The strategic lesson is architectural modularity: model labs may compete less through a single branded attention doctrine and more through combinations tuned to workload, memory, and serving cost.
  • Sparse attention's ability to preserve important long-range information will determine whether this convergence is a real replacement for global attention or merely a cost compromise.
Open at Leiphone Ask Claude ChatGPT

Beijing adds incremental support while sharpening its external message

# Policy Sinocism 22h

A China-focused policy roundup says Beijing is adjusting monetary tools, managing Japan ties, and weighing cooperation on AI governance, showing incremental rather than stimulus-heavy intervention.

More
  • The item says the State Council called for practical additional policies, followed by central-bank adjustments to structural monetary tools and finance-ministry support measures.
  • Its framing presents the policy response as incremental management of growth pressure rather than a single large rescue package.
  • The roundup also flags a message to Japan and sensitivities around Xi's Washington visit, linking domestic policy to foreign-policy signaling.
  • The mention of possible US-China work on AI or strategic-intelligence governance is not detailed enough to establish an agreement, but it points to governance as a new bilateral channel.
Open at Sinocism Ask Claude ChatGPT

Step 5 Preview returns to the open-model frontier

# AI Leiphone 12h

Chinese lab Step says Step 5 Preview ranks among the world's top open models and can build a working 3D game across Blender and web tooling, though long-task reliability still needs proof.

More
  • The model reportedly scored 44 on Artificial Analysis's intelligence index, placing it second among open models in the supplied account.
  • A hands-on test asked it to create original assets in Blender, export GLB files, build a Three.js game, implement movement and collisions, and run tests and a production build.
  • The task was divided into stages, but the model itself handled file creation, tool use, debugging, and handoff without the authors writing the code.
  • The resulting game was playable and structurally complete, though the article acknowledges rough visual and interaction details.
  • The important capability is sustained cross-tool execution: the model can preserve a project goal while moving between a 3D editor, asset pipeline, and web application.
  • The remaining question is whether this reliability survives broader tasks and independent evaluation rather than curated demonstrations.
Open at Leiphone Ask Claude ChatGPT

Chinese factories are turning industrial AI into an OEE problem

# Analysis QbitAI 8h

Chinese industrial AI operators say factory systems must improve measurable equipment efficiency and survive site variation, not merely score well on model benchmarks.

More
  • A semiconductor wafer-cutting example uses multimodal recognition and agentized operator knowledge; the vendor reports about a 25% increase in equipment effectiveness and capacity.
  • Industrial deployments face sparse, confidential, and site-specific data, so a model that works in one plant may need substantial adaptation in another.
  • The proposed scaling path is an ecosystem: a platform connects industrial AI vendors to factories, systems integrators, and delivery partners.
  • The speakers argue that high-value applications with ready data and measurable effects on yield, process control, or OEE will mature before general-purpose factory agents.
  • The China-specific signal is organizational as much as technical: platform companies are trying to turn isolated proofs of concept into repeatable industrial delivery.
Open at QbitAI Ask Claude ChatGPT

A Chinese scholar reads the Xi-Trump summit as strategic stabilization

# Analysis Ginger River Review 7h

Fudan scholar Wu Xinbo presents the summit as progress toward constructive strategic stability, reflecting how Chinese official and academic voices frame managed competition.

Also: SCMP

More
  • The interview says China and the United States should build relations around respect, fairness, reciprocity, and regularized engagement.
  • It highlights an alleged arrangement to reduce tariffs on about $30 billion of goods from each side, with most covered products moving toward most-favored-nation rates.
  • This is an interpretation from a prominent Chinese foreign-policy scholar, not an independent account of what Washington agreed to.
  • The framing prioritizes stability and reciprocal treatment, while leaving asymmetric concerns and enforcement details less prominent than they would be in much US coverage.
  • The contrast is useful because it shows Beijing presenting summit diplomacy as institutionalizing competition rather than resolving the underlying rivalry.
Open at Ginger River Review Ask Claude ChatGPT

Chinese office agents still leave humans with the liability

# Analysis Leiphone 13h

A hands-on comparison of WorkBuddy, Qwen Office, and Doubao Work found all three could produce deliverables, but their quarterly totals differed by 57.7%, making verification the real bottleneck.

More
  • The test used the same messy sales ledger with inconsistent dates, refunds, missing departments, and anomalous amounts, then asked each agent for spreadsheets, slides, and a summary.
  • The highest and lowest quarterly totals reportedly differed by 57.7%, showing that document production can conceal divergent business judgments.
  • The article characterizes the products as different bets: Tencent on a platform, Alibaba on an organizational entry point, and ByteDance on enterprise context.
  • Their large desktop clients suggest these are execution environments with files, credentials, browsers, and permissions rather than simple chat windows.
  • The core management problem is accountability: agents can perform actions, but people still own reconciliation, interpretation, and sign-off.
  • This is a valuable China-specific benchmark because it tests office agents against messy local work rather than polished demo tasks.
Open at Leiphone Ask Claude ChatGPT

China adds a 55% tariff on Brazilian beef after the quota fills

# Policy SCMP 2h

China imposed an extra 55% tariff on Brazilian beef after Brazil exhausted its 1.106-million-ton quota, rejecting a request to use unused Uruguayan quota.

More
  • The move shows Beijing using tariff-rate quotas as an active bargaining instrument rather than a purely administrative import limit.
  • The source says the Ministry of Commerce announced the surcharge after customs confirmed the Brazilian quota was full.
  • The decision also illustrates how agricultural trade disputes can spread across neighboring countries when quota allocation is not transferable.
  • For exporters, the strategic risk is not only the headline tariff but the possibility that Beijing will enforce origin-specific limits even when regional capacity remains unused.
Open at SCMP Ask Claude ChatGPT

Xi calls for confidence and stronger growth in National Day speech

# Policy SCMP 6h

Xi Jinping's National Day address combines confidence in overcoming external pressure with a call for robust growth and greater certainty, signaling the official economic message.

More
  • The speech was delivered at a state reception marking the 77th anniversary of the People's Republic and followed Xi's recent US visit.
  • The supplied account emphasizes confidence and economic momentum rather than announcing a new policy package.
  • That framing matters because Beijing is trying to project stability while households, firms, and foreign investors continue to look for clearer growth signals.
  • The address also places geopolitical complications inside a domestic mobilization narrative: challenges are acknowledged, but the prescribed response is confidence and continued development.
Open at SCMP Ask Claude ChatGPT

The EU prepares new trade tools against China

# Policy SCMP 12h

Brussels and Berlin are hardening their China stance before a decisive October, with new trade instruments likely to arrive in December and senior talks scheduled in Beijing.

More
  • The European trade chief is due to meet China's commerce minister in Beijing on October 8 and 9, making the visit a near-term test of whether dialogue can slow escalation.
  • The report describes a pipeline of tools aimed at Chinese trade practices rather than one isolated tariff or investigation.
  • The China dimension is substantive and bilateral: European policy is shifting from market-access complaints toward a broader ability to retaliate or constrain exposure.
  • For Chinese exporters and foreign companies operating in Europe, the risk is cumulative regulatory friction that arrives through several instruments at once.
  • The article gives no final measure or implementation text, so the significance lies in direction and timing rather than a completed policy.
Open at SCMP Ask Claude ChatGPT

Personal agents are moving from task tools to persistent digital identities

# AI Leiphone 12h

Chinese coverage sees Muse, Dot, Cue, and domestic projects competing for a persistent context layer that can act across browsers, mail, phones, wallets, and platforms.

More
  • Office agents manage a bounded task, while personal agents try to maintain a durable model of a person's preferences, history, and unfinished goals.
  • The architecture increasingly pairs long-term memory with a cloud computer, browser, communications identity, and sometimes payment capability, making authorization and auditability central.
  • The strategic battle is for the pre-platform decision point: a user could ask an agent to buy or book something before ever opening a marketplace or search engine.
  • Chinese firms including Qwen, Doubao, Tencent, and ByteDance are reported to be pursuing similar products, showing that the pattern is entering China's domestic product roadmap.
  • The hardest engineering problem is relevance over time: deciding which memories are current and task-relevant rather than merely available.
  • This is a platform shift, not just a new assistant category, because the agent could mediate traffic, transactions, and user intent before incumbent services see the request.
Open at Leiphone Ask Claude ChatGPT

Did China and the US actually agree on crisis prevention?

# Policy SCMP 8h

The summit's Chinese and US accounts describe military crisis-prevention talks differently, exposing how wording and implementation can diverge even when leaders claim progress.

More
  • Xi publicly called for regular military dialogue and better crisis-communication mechanisms, while China's post-summit statement said the two sides agreed to conclude a memorandum.
  • The item says analysts see the disagreement in wording as a possible reason the two governments characterized the outcome differently.
  • For operators and policymakers, the important risk is not whether a phrase appears in a communiqué but whether hotlines, exercises, and incident protocols gain usable authority.
  • Chinese framing foregrounds a negotiated institutional commitment, while US-facing interpretation appears more cautious about whether anything binding was actually agreed.
  • The episode is a reminder that summit language is itself contested infrastructure: ambiguity can preserve diplomatic space but also limit deterrence and crisis-management value.
Open at SCMP Ask Claude ChatGPT

Li Auto wants battery suppliers to share the whole data stack

# Policy 36Kr 15h

Li Auto says any cell maker may work with it only inside Li's battery architecture, forcing suppliers to share production data and accept the automaker's system-level control.

More
  • Li Auto designs the cells and pack while suppliers manufacture cells; it also keeps control of battery management, thermal management, insulation, and integration.
  • The company says it has moved away from a backup-supplier model and will require every partner to meet its own technical standards regardless of price.
  • The strategy has already reduced CATL's position: the article says Li left CATL's top-five customer list in August and no longer includes it in the current partner lineup.
  • Li argues that end-to-end data lets it detect manufacturing defects earlier, coordinate battery, thermal, drive, and charging systems, and own the customer warranty experience.
  • The company reports more than 300 battery engineers, more than 2 billion yuan, roughly $280 million, in annual research spending, and cumulative investment of over 10 billion yuan, roughly $1.4 billion.
  • This is a supply-chain power shift: automakers are trying to commoditize cell manufacturing while retaining the data, definition, and liability layers that determine vehicle safety and differentiation.
Open at 36Kr Ask Claude ChatGPT

China tells microdrama makers to slow down and protect rights

# Policy SCMP 16h

China's broadcasting regulator is pushing live-action, higher-quality microdramas and stronger copyright safeguards as AI lowers production costs and increases low-quality output.

More
  • The policy message treats AI as both a production tool and a source of content-quality and rights risks in the fast-growing microdrama market.
  • By favoring live-action and quality controls, regulators are steering an industry built on rapid, inexpensive production toward more conventional editorial and licensing norms.
  • The move suggests that China's media governance is responding not only to what content says but to how generative tools reshape the economics of production and imitation.
  • For creators and platforms, compliance may increasingly include provenance, rights clearance, and limits on synthetic or derivative material rather than only content review.
Open at SCMP Ask Claude ChatGPT

China launches mortgage subsidies as AI and chip investment accelerate

# Policy 36Kr 20h

A Chinese news roundup leads with one-percentage-point mortgage subsidies for qualifying homes, while also noting a major memory investment plan and rapid AI product expansion.

More
  • From October 1, eligible new commercial home loans can receive a one-percentage-point annual subsidy for up to five years, with a reported maximum benefit near 50,000 yuan, roughly $7,000.
  • The subsidy targets first-home buyers purchasing homes up to 120 square meters and 1.5 million yuan, roughly $211,000, signaling support for household demand rather than broad property rescue.
  • The same roundup says ChangXin plans 34.9 billion yuan, roughly $4.9 billion, in new projects, illustrating continued state-linked capital formation in memory.
  • It also frames OpenAI's personal agent and cheaper model as part of a broader global shift toward always-on AI, but those foreign items are secondary here.
  • The value of the roundup is the policy mix: Beijing is trying incremental demand support while continuing strategic industrial investment, rather than choosing between property stabilization and technology spending.
Open at 36Kr Ask Claude ChatGPT

Chinese researchers model lidar detection of an F-35 at 61.5 kilometers

# Analysis SCMP 10h

A Chinese simulation claims a missile-borne lidar could detect and track an F-35 at up to 61.5 kilometers at night, but the result is theoretical and contested.

More
  • The source attributes the model to a leading Chinese air-force missile supplier and says the range applies under ideal conditions.
  • A lidar-based seeker would matter only if atmospheric attenuation, target signature, pointing, processing, and missile kinematics all support the modeled result.
  • The item includes analyst skepticism, which is important because a simulation range is not an operational detection capability.
  • If validated, the work would challenge assumptions about the survivability of stealth aircraft in specific nighttime and engagement geometries, not eliminate stealth generally.
  • The story is valuable as a window into Chinese defense research priorities, but the supplied evidence does not justify treating 61.5 kilometers as a fielded capability.
Open at SCMP Ask Claude ChatGPT

China's AI hardware bottleneck is increasingly a scheduling problem

# Signal InfoQ China 3h

An IDC assessment argues that fragmented chips, whole-card allocation, and poor gang scheduling leave much of China's purchased AI capacity unusable, especially for bursty agent workloads.

More
  • The report identifies three failure modes: incompatible software stacks across accelerators, coarse whole-card allocation, and distributed jobs waiting for incomplete resource sets.
  • Agent workloads make the problem harder because long tool loops alternate between short compute bursts and idle periods, creating demand tides that static reservations handle badly.
  • The proposed architecture combines fine-grained sharing with isolation, gang scheduling for distributed jobs, and hardware-neutral pools for exclusive, shared, and oversubscribed use.
  • The item is tied to a vendor's own IDC ranking, so its claims about the leading provider should be treated as commercial positioning rather than independent proof.
  • The strategic lesson is that China's accelerator buildout will not translate into service capacity without a control plane that measures token throughput, latency, and cost rather than installed cards.
Open at InfoQ China Ask Claude ChatGPT

Chinese energy group claims a hydrogen-boron fusion milestone

# Signal Solidot 5h

Newao says its Xuanlong-50U device repeatedly measured hydrogen-boron reaction products above 100 million reactions per second, a promising but company-reported fusion claim.

More
  • The company says it combined high-energy neutral-beam injection with radio-frequency heating to increase the fast-proton share near hydrogen-boron resonance.
  • Hydrogen-boron fusion is attractive because it could produce mainly alpha particles, but it requires much harsher conditions than deuterium-tritium fusion.
  • The source says outside fusion experts reviewed the experiment and considered the particle and alpha-product measurements repeatable, but it provides no peer-reviewed paper or independent dataset.
  • The result should therefore be read as a diagnostic milestone, not evidence of net energy gain or a practical reactor.
  • Its strategic value is that Chinese fusion programs continue exploring multiple fuel cycles rather than converging exclusively on the mainstream deuterium-tritium route.
Open at Solidot Ask Claude ChatGPT

Cambricon wins a long-running executive equity dispute

# Policy Leiphone 6h

Cambricon says former executive Liang Jun lost six challenges over an equity buyback and is now subject to enforcement, highlighting the hard edge of Chinese retention schemes.

More
  • The company's filing says leaving during the lock-up period triggered a contractual repurchase clause and that final judgments upheld the arrangement.
  • Liang reportedly increased his damages claim, but Cambricon says the dispute does not affect daily research or operations.
  • The case shows how Chinese technology companies use partnership and equity plans to bind senior technical staff beyond ordinary employment terms.
  • It also illustrates the asymmetry between contractual incentives and employee liquidity: a departing executive can lose the expected upside while still facing enforcement costs.
  • The source presents Cambricon's legal position and does not include Liang's response, so the factual record should be read as one-sided.
Open at Leiphone Ask Claude ChatGPT

China's non-compete agreements are trapping ordinary workers

# Culture TMTPost 9h

Chinese employees describe six- and seven-figure yuan claims after leaving large technology companies, exposing how non-competes can become low-cost retention and labor-control tools.

More
  • One former e-commerce worker reportedly faced a 1.47 million yuan penalty, roughly $207,000, plus 103,600 yuan, roughly $14,600, in returned compensation after an absentee arbitration ruling.
  • The article says an opposition network has gathered more than 3,000 workers and argues that large companies use broad competitor lists and aggressive claims beyond genuinely sensitive roles.
  • The legal mechanism can make a worker unemployable across an industry while litigation or arbitration proceeds, shifting the balance of risk sharply toward employees.
  • The case also shows how surveillance and procedural opacity matter: the worker says tracking evidence and notices were misunderstood or never effectively received.
  • For technology executives, the issue is not only compliance but labor-market design; overly broad restrictions can preserve secrets while damaging mobility, trust, and the supply of experienced engineers.
  • The source is a reported labor account with disputed facts, but the scale and recurring company names suggest a structural governance question rather than one isolated dispute.
Open at TMTPost Ask Claude ChatGPT

BYD reaches 2,000 highway fast-charging stations

# Signal Leiphone 12h

BYD says its highway network now spans 2,000 stations and can charge a car from 10% to 97% in about nine minutes, but the holiday test is really about queueing and grid operations.

More
  • The network reportedly covers 278 highways, including major north-south and east-west corridors, and is open to vehicles from other brands.
  • Fast single-vehicle charging does not guarantee short waits; simultaneous demand, bay turnover, grid capacity, and fault response determine the actual travel experience.
  • BYD is using cameras, voice prompts, storage, grid coordination, and on-site staff to prevent noncharging vehicles from occupying charging bays.
  • The company's high-altitude station near Everest shows how infrastructure can create demand by making previously difficult destinations viable for electric vehicles.
  • The source says BYD aims for 20,000 total fast-charging stations by year-end, making operational reliability and capital efficiency more important than the headline count.
  • This is a useful China infrastructure signal: automakers are building proprietary energy networks that also function as product differentiation and a data loop.
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 1/20
Pekingnology 200 1/20
Ginger River Review 200 1/20
Sinocism 200 1/20
QbitAI 200 9/10
36Kr 200 6/30
Solidot 200 12/20
Leiphone 200 20/20
GeekPark 200 1/30
IT Home 200 12/60 +39 over cap
SCMP 200 10/50 +6 over cap
Initium 200 4/15
Jiemian ERR 0/0 ConnectionError: HTTPSConnectionPool(host='a.jiemian.com', port=443): Max retries exceeded with url: /index.php?m=article&a=rss (Caused by NewConnectionError("HTTPSConnection(host='a.jiemian.com', por
China Digital Times 200 3/15
The Wire China 200 0/50
Techmeme 200 1/15
Hacker News 200 0/20
V2EX 200 8/40 +32 over cap
Zhidx 200 7/20
InfoQ China 200 8/20 +5 over cap
New Wisdom 200 8/15
TMTPost 200 8/19 +11 over cap