China Desk

China News, Summarized 01 Oct 2026, 05:39 UTC 34 stories 22/23 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 details the sandbox platform behind its agent training

# QbitAI 1d

DeepSeek and Tsinghua describe DSec, a stateful sandbox system supporting model training, evaluation and data preparation across containers, microVMs and full virtual machines.

More
  • The system matches execution back ends to task risk: reused containers for short evaluations, containers for general engineering, microVMs for security work and full VMs for graphics or Android.
  • DSec separates base images, workspaces and toolkits so each layer can be versioned and composed without rebuilding everything.
  • In a test creating 8,192 containers, on-demand EROFS loading cut completion time from more than an hour to about 35 minutes and reduced disk writes 57%.
  • Around 90% of sandboxes reportedly used less than 5% of their requested CPU on average while retaining memory, enabling more than 50-fold CPU oversubscription.
  • The architecture shows that agent training depends on an operating system for stateful experimentation, not merely a GPU scheduler or a model API.
Open at QbitAI Ask Claude ChatGPT

Anthropic warns that an open Chinese model can exploit vulnerabilities

# QbitAI 19h

Anthropic's report says Zhipu's GLM-5.3 performed near its own cyber model on exploit tests and that its safeguards were easier to weaken, while Chinese commentary calls the warning an advertisement.

More
  • Anthropic reportedly saw GLM-5.3 succeed 50 times in 410 exploit attempts versus 56 for its own preview model, a strikingly narrow gap in the cited test.
  • A smaller version allegedly built a stable ARM64 exploit chain around public Chrome flaws after several hours of autonomous work, at an estimated API cost of $20.40.
  • The report says modifying open weights sharply reduced refusal rates, but the Chinese analysis notes that this is a general property of open-weight models rather than a unique flaw.
  • The dispute exposes a policy split: closed models can hide weights and control interfaces, while open models spread capability and make safety layers easier to remove.
  • The strongest signal is capability diffusion. Cyber offense that was once restricted to frontier labs is increasingly appearing in open models, forcing independent testing and controlled access.
Open at QbitAI Ask Claude ChatGPT

DeepSeek gives Huawei Ascend a CUDA alternative at the kernel layer

# Leiphone 18h

DeepSeek's Ascend port of TileLang and five core libraries is framed in Chinese coverage as a bid to loosen CUDA dependence through open operator infrastructure.

Also: InfoQ China

More
  • TileLang abstracts tiled computation while still exposing hardware-specific scheduling, memory movement and synchronization, aiming to combine portability with near-native performance.
  • The port covers GEMM, MoE communication, sparse attention, common kernels and top-k selection, which are central to training and inference cost.
  • The strategic value is not a single benchmark but a shared development path: one front end can target Nvidia and Ascend while preserving model-level interfaces where possible.
  • The report claims large speedups and cross-platform performance, but the accompanying evidence is vendor or article benchmarking rather than an independent test.
  • Open kernel ecosystems can matter more than a single model release because they determine whether developers can continuously tune workloads on domestic hardware.
Open at Leiphone Ask Claude ChatGPT

Personal agents are moving between the user and the internet

# Leiphone 21h

Chinese analysis says OpenAI's Dot, Meta's Muse, Manus Cue and domestic products are competing to become the first interface users consult before apps, search or commerce.

Also: TMTPost

More
  • The core product shift is from completing a bounded task to maintaining a model of a person's preferences, accounts, schedules and unfinished intentions.
  • Cloud computers, browsers, phone numbers, wallets and persistent memory give agents the ability to act across services rather than merely recommend actions.
  • The Chinese angle is especially important because local agents must connect to domestic messaging, office and payment systems while obeying data rules.
  • This is an entry-point battle: whoever receives the user's natural-language request can mediate search, advertising, transactions and platform choice.
  • The unresolved constraints are identity, permissions, payment risk, stale memory and the economics of running an always-on assistant.
Open at Leiphone Ask Claude ChatGPT

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

# Leiphone 21h

Huawei says 5,500 Hi-One optical engines could replace 48,000 pluggable 800G modules in a 4,096-card Ascend 960 supernode, shortening electrical links to about five centimeters.

More
  • Near-packaged optics sits between pluggable modules and co-packaged optics, aiming to reduce signal loss without making field maintenance as difficult as fully integrated designs.
  • Each Hi-One is described as having 36 lanes at 200G for 7.2 terabits per second, with a shared multi-wavelength light source and silicon photonics.
  • The article highlights unresolved engineering questions about laser redundancy, channel allocation, thermal behavior and whether all 36 lanes carry normal traffic.
  • Huawei's scale claims suggest that optical interconnect, not only NPU compute, is becoming a differentiator in domestic supernode architecture.
  • The public parameters are insufficient to reconstruct the real topology, so the most valuable insight is the design tradeoff rather than the headline bandwidth.
Open at Leiphone Ask Claude ChatGPT

CNNIC says more than 700 million Chinese use generative AI

# Solidot 14h

CNNIC says China's generative-AI user base exceeded 700 million in the first half of 2026, with question answering dominant and office and media tasks spreading.

More
  • The reported penetration exceeds half of China's internet population, making generative AI a mass consumer service rather than a specialist tool.
  • CNNIC says 76% of users ask questions, while 47.8% use AI for image or video work and 32.5% for summaries, meeting notes or presentations.
  • The report also says 38.7% of internet users bought smart hardware recently, linking AI adoption to device purchasing rather than software alone.
  • The numbers are from a state-affiliated survey and should be read with attention to sampling and definitions; they are still a useful measure of official market framing.
  • The claimed dominance of Chinese models in global call rankings is not independently substantiated in the supplied report.
Open at Solidot Ask Claude ChatGPT

Chinese research vessel near Philippine islands raises cable concerns

# SCMP 1h

A Chinese civilian research ship spent more than a week near the Philippines' Batanes Islands, intensifying concern over maritime surveillance and undersea cables.

More
  • The vessel's AIS track places a nominally civilian platform in a strategically sensitive area near routes and waters contested by Beijing and Manila.
  • The report connects the activity to wider Chinese pressure against Philippine and Japanese maritime arrangements, but does not establish cable interference.
  • The episode illustrates how research, mapping and civilian maritime operations can create gray-zone uncertainty without a formal military deployment.
  • For infrastructure operators, the strategic issue is the difficulty of distinguishing benign ocean science from data collection relevant to future disruption.
Open at SCMP Ask Claude ChatGPT

Chinese envoy calls Xi-Trump rapport a strategic asset

# SCMP 3h

Ambassador Xie Feng described the personal relationship between Xi Jinping and Donald Trump as the most valuable stabilizing asset in bilateral ties.

More
  • The wording elevates leader-to-leader rapport above institutional mechanisms, signaling Beijing's preference for summit diplomacy as a risk-control channel.
  • The statement came at a National Day event in Washington, so it also functions as diplomatic messaging to an American audience.
  • Personal rapport can create room for trade or crisis bargains, but it is inherently vulnerable to domestic politics and sudden policy shifts.
  • The framing shows how Chinese officials present stability as a product of top-level trust rather than convergent strategic interests.
Open at SCMP Ask Claude ChatGPT

China plans to shift consumption-tax collection toward local governments

# Jiemian 37m

Finance Minister Lan Foan called for moving consumption-tax collection downstream and allocating more revenue to local governments while tightening debt discipline.

More
  • The proposal addresses a structural mismatch: local governments carry spending obligations but have limited own-source revenue after the property downturn.
  • The same statement calls for resolving hidden debt, ending new financing-platform borrowing and protecting basic public services.
  • Moving consumption-tax revenue to localities would change incentives around retail activity and regional tax competition, but the supplied report gives no timetable or final design.
  • The policy is part of a broader attempt to make local finances less dependent on land sales and opaque borrowing.
Open at Jiemian Ask Claude ChatGPT

Huawei unveils new Kirin chips for the Mate 90 family

# IT Home 42m

Huawei says the Mate 90 line uses new Kirin 9030, 9035, 9050 and 9050 Pro chips, with claimed gains in CPU, GPU and neural processing.

More
  • Huawei divides the chips into flagship and so-called folded-logic tiers, suggesting a product segmentation strategy built around domestic silicon rather than a single flagship die.
  • The company claims the 9050 Pro raises NPU performance 140% over its predecessor, while the supplied report does not provide process node, packaging or independent benchmarks.
  • The chips are paired with Huawei's hardware, software and cloud stack, reinforcing vertical integration as the main strategic response to external technology restrictions.
  • The launch is important less for the marketing names than for Huawei's continued ability to ship a high-end domestic silicon platform.
Open at IT Home Ask Claude ChatGPT

Chinese industrial robot maker grows abroad while losses persist

# TMTPost 4h

Stander Robotics reported 139% year-on-year revenue growth in the first four months of 2026, but still lost 61.84 million yuan, roughly $8.7 million, as overseas orders drove expansion.

More
  • Foreign revenue reached 67.9% of the period total, driven mainly by large orders from system integrators in Japan, Taiwan and Thailand.
  • The business began with autonomous mobile robots for factory logistics and is extending that engineering base toward embodied robots and software.
  • Overseas gross margin was higher than domestic margin, but it declined as the company used more competitive pricing to win orders.
  • The dependence on integrators makes expansion faster but leaves Stander exposed to project timing, local service capability and repeat-purchase risk.
  • The third Hong Kong listing attempt will test whether investors value export growth before the new embodied-robot business produces durable profit.
Open at TMTPost Ask Claude ChatGPT

Tencent reportedly signs a $7 billion Oracle deal for overseas AI chips

# Techmeme 49m

Sources cited by the Financial Times say Tencent agreed to spend about $7 billion over five years for access to roughly 100,000 advanced chips through Southeast Asian data centers.

More
  • The reported structure would put restricted accelerators outside China while still serving a Chinese platform company, making data-center geography part of the supply-chain strategy.
  • The scale suggests Tencent is competing with ByteDance and Alibaba not only for models but for guaranteed inference capacity.
  • Because the item is an aggregator summary of unnamed-source reporting, the contract terms and chip count remain unverified here.
  • If confirmed, the deal would show export controls shifting cloud architecture and capital spending offshore rather than eliminating Chinese demand.
Open at Techmeme Ask Claude ChatGPT

Chinese coverage highlights Gemini 4 Argon's benchmarks and caveats

# AI Zhidx 3h

Chinese tech coverage presents Gemini 4 Argon as a major enterprise-model challenger while noting reports that some Google employees found coding performance less consistent in practice.

Also: InfoQ China

More
  • The coverage emphasizes the 1-million-token output ceiling, long-horizon software engineering and enterprise benchmarks, including 77.9% on DeepSWE.
  • It also reports lower cost claims and a 95% cache discount, but these figures depend on Google's pricing and benchmark methodology.
  • The most useful caveat is the contrast between public scores and reported internal experience, a familiar warning that long-task benchmarks can overstate reliability.
  • Chinese coverage treats Argon primarily as evidence that the frontier is moving toward agentic enterprise execution rather than ordinary chat.
Open at Zhidx Ask Claude ChatGPT

GPT-6 controls a Unitree G1 through a kitchen skill library

# AI QbitAI 21h

A Stanford project lets GPT-6 Astra explore an unfamiliar kitchen, remember object locations and call navigation, grasping and drawer-opening skills on a Unitree G1.

More
  • The system separates high-level planning from low-level control: the model chooses a target and skill, while perception, motion planning and balance controllers execute the movement.
  • The robot builds spatial memory from camera, lidar and pose data, then aligns it with an Isaac Sim digital kitchen for planning and retrieval.
  • Astra can combine tasks such as collecting trash and retrieving medicine from a drawer even when the object is not currently visible.
  • No new data collection or task-specific action-policy training was reportedly required for the new kitchen, although the underlying perception and control models were pretrained.
  • The architectural lesson is that general-purpose models become more useful when they call well-defined skills instead of trying to generate motor commands directly.
Open at QbitAI Ask Claude ChatGPT

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

# Analysis ChinaTalk 19h

An interview with Rhodium's Logan Wright argues that China's property collapse, weak demand, debt trap and export push are facets of one financial model that now constrains growth.

More
  • The argument is not that China has no choices, but that every viable option implies slower near-term growth and fiscal and financial restructuring.
  • The old engine relied on credit-funded investment and property; diminishing returns now make more of the same less effective while local governments face revenue constraints.
  • China's push into EVs, AI and robotics can create industrial strength but cannot quickly replace the employment and demand once supplied by property and construction.
  • The labor challenge is especially acute with 12 million to 13 million graduates entering the workforce each year, while ideology and fiscal limits narrow policy options.
  • The interview's strategic conclusion is that competition is shifting from a pure growth contest toward military capability and industrial resilience.
Open at ChinaTalk Ask Claude ChatGPT

Robot listings force China's market to ask who is really selling

# Analysis QbitAI 1d

A survey of 36 listed Chinese robotics companies finds a widening gap between genuine robot revenue, component exposure and speculative humanoid positioning.

More
  • The sample splits evenly between companies selling complete robots or applications and those supplying systems or components, making the sector much broader than humanoid startups.
  • Unitree reportedly derived more than half of 2025 revenue from humanoid robots, while UBTech's humanoid share was about 41%; other companies remained in single digits.
  • The emerging regulatory test is whether revenue is real, recurring and replicable, not whether a company has an impressive demonstration or partnership announcement.
  • The comparison also separates businesses with established industrial, warehouse or household deployments from companies whose embodied-AI revenue is still experimental.
  • The shift matters for executives because commercial proof is moving from demos and funding rounds into audited product mix, repeat orders and cash generation.
Open at QbitAI Ask Claude ChatGPT

Manus returns with an agent that has a phone and wallet

# AI QbitAI 21h

Manus 2.0 adds a Studio workbench and Cue, a personal agent with its own email, phone number, wallet and computer that can collaborate with other agents.

More
  • Cue is designed as an identity and execution bundle: it can contact people, make authorized payments, take calls and leave summaries.
  • Multiple agents can be placed in one group for a shared project, turning delegation into a visible coordination model rather than a hidden chain of tool calls.
  • Manus Studio adds editable video timelines, game development and persistent cloud computers, keeping human intervention near the final artifact.
  • The company says its Cascade framework cut token use 23.2%, task time 28.2% and operating cost 32.0% in internal tests.
  • Manus is returning to a crowded field after its Meta acquisition and spinout, so the strategic question is whether its integrated infrastructure can outlast the novelty of agent identities.
Open at QbitAI Ask Claude ChatGPT

Chinese engineers attack Kimi K3's latency with fused kernels

# AI Zhidx 19h

Chinese and foreign teams are optimizing Kimi K3 by fusing operators, communication and memory movement, showing that trillion-parameter inference is becoming a systems problem.

More
  • The model's 2.8 trillion total parameters and 104 billion active parameters create bandwidth and all-to-all communication bottlenecks that raw FLOPS cannot solve.
  • One approach puts all 92 MoE layers into a single TPU kernel with asynchronous weight movement; another fuses many operators into larger kernels on a tightly coupled domestic-chip system.
  • The shared principle is to remove launch boundaries and keep intermediate data on-chip while overlapping computation, communication and prefetching.
  • The article says an AI agent generated and refined super-operators for Kimi K3, suggesting models may increasingly optimize their own inference substrate.
  • The reported speedups are platform-specific and should not be generalized across chips or workloads.
Open at Zhidx Ask Claude ChatGPT

Chinese AI infrastructure firm files for a Hong Kong listing

# Signal Zhidx 1d

Beijing AI infrastructure provider Jiuzhang Yunji filed for a Hong Kong IPO after 2025 revenue reached 1.098 billion yuan, roughly $155 million, up 177.7%.

More
  • The company is shifting from project-based data-center integration toward a token factory and managed inference service, hoping to capture recurring usage revenue.
  • It reported 29,000-plus PFLOPS of controllable capacity across seven provinces and 79.6% cloud-resource utilization in the first half of 2026.
  • The business remains financially fragile: cumulative losses were about 1.188 billion yuan, roughly $167 million, through 2025, while first-half 2026 reported a 1.095-billion-yuan loss largely from fair-value changes.
  • Infrastructure solutions still generated 73.8% of first-half revenue, so the token-service strategy is not yet the dominant business.
  • The filing illustrates China's effort to create AI-native cloud providers between hyperscalers and hardware integrators, with state-linked capital visible among earlier investors.
Open at Zhidx Ask Claude ChatGPT

OpenAI Dot is presented as a chief of staff for personal work

# AI Zhidx 17h

Chinese coverage of Dot highlights its cloud computer, long-running tasks and ability to delegate work across threads, distinguishing it from consumer assistants.

Also: New Wisdom

More
  • A reported demonstration had Dot audit sponsorship records, find missing invoices and identify a $10,000 discrepancy, illustrating the value of persistent retrieval rather than one-shot answers.
  • The proposed product pattern is a coordinator that reads context, breaks goals into subtasks and delegates coding, research or image work to specialized threads.
  • The Chinese framing compares Dot with domestic efforts by Tencent, Alibaba and ByteDance, where local files, messaging and enterprise permissions are central differentiators.
  • The hard product question is not whether an agent can produce artifacts but who verifies the result and controls the accounts it can access.
Open at Zhidx Ask Claude ChatGPT

Kuaishou reorganizes around enterprise AI and Kling

# Policy Zhidx 18h

Kuaishou's founder will oversee the community science line while its former head becomes Kling AI chief, alongside a new cross-company enterprise-AI productivity group.

More
  • The new group combines workplace infrastructure, general agents, enterprise context and internal information systems, signaling that Kuaishou treats AI as organizational plumbing rather than a standalone product.
  • Moving a senior technical leader to Kling gives the video model a clearer operating owner while the founder takes direct responsibility for a core recommendation and community science function.
  • The arrangement reflects a broader Chinese internet-company pattern: centralize data and agent infrastructure, then assign business leaders to turn models into products.
  • Kuaishou says AI is becoming the company's engine across recommendation, search, ads, content and commerce, but the announcement gives no financial targets.
Open at Zhidx Ask Claude ChatGPT

Industrial AI in China moves from pilots toward repeatable delivery

# Analysis QbitAI 17h

A Chinese discussion with Siemens and Aqiu Technology argues that factory AI must be measurable, reliable and repeatable across sites rather than merely impressive in a demo.

More
  • A semiconductor case reportedly improved equipment effectiveness and capacity by about 25% after multimodal perception and agentic troubleshooting were introduced.
  • Industrial systems are difficult to copy because factories differ in equipment, data, process rules and safety requirements even within one company.
  • The proposed platform role is to connect validated technology with customers, integrators and delivery partners rather than pretend one vendor can standardize every factory.
  • The strongest initial opportunities are high-value applications around yield, process control and equipment effectiveness where data is available and errors are expensive.
  • The discussion frames industrial AI as a systems-integration and commercial-repeatability problem, not a model leaderboard problem.
Open at QbitAI Ask Claude ChatGPT

A Chinese robot learns Suzhou embroidery through touch and force

# AI Zhidx 1d

AGILINK's 21-degree-of-freedom OmniHand 3 Ultra is being trained on delicate embroidery tasks involving needles, thread, silk and human collaboration.

More
  • Embroidery combines deformable materials, submillimeter alignment, multi-finger coordination, force control and continuous state transitions, making it a tougher test than stable-object grasping.
  • The hand uses fingertip visual-tactile sensors and tactile coverage across the fingers and palm to estimate contact and grip conditions.
  • The reported needle eye is only 0.15 millimeters wide, while later thread-splitting and stitching stages require adapting to tension changes rather than following a fixed trajectory.
  • The system completed early thread-splitting steps independently and used demonstration, teleoperation data and reinforcement learning for later operations.
  • The cultural task is valuable as a benchmark: skills learned on silk and thread could transfer to wiring, clothing, assembly and other flexible industrial materials.
Open at Zhidx Ask Claude ChatGPT

GPT-6 can build a digital Manhattan but still needs a manager

# AI Leiphone 21h

Early tests suggest GPT-6 Astra can operate Blender, Final Cut and Unreal Engine for extended tasks, but long projects still require human direction and stopping decisions.

More
  • The model reportedly checks rendered pages, inspects errors and revises results, moving beyond one-pass generation toward a test-and-repair loop.
  • Users demonstrated static 3D assets, browser games and video-editing workflows, indicating that computer control is expanding beyond code and text.
  • In a week-long Manhattan build, Astra reportedly over-optimized individual buildings while leaving broader work incomplete, showing local competence without project-level prioritization.
  • The human tester acted as a manager loop, deciding when detail was sufficient, what to do next and whether the project remained on course.
  • The strategic boundary is clear: stronger agents can sustain execution longer, but they still do not reliably define a useful endpoint for ambiguous work.
Open at Leiphone Ask Claude ChatGPT

OpenAI researcher says multi-agent scale is still poorly understood

# Analysis QbitAI 15h

Noam Brown says 10,000 agents contributed at most 10% to a reported mathematics breakthrough, arguing that model strength mattered more than agent count.

More
  • Brown describes multi-agent systems as parallel test-time computation: useful when a problem decomposes, but less efficient than one agent with full context.
  • He says small experiments with four or sixteen agents show gains, while extrapolating to 10,000 agents is too expensive for rigorous ablation studies.
  • The interview distinguishes a compelling demonstration from a scientific scaling law; the large token count and short elapsed time do not prove proportional capability.
  • The Chinese report is valuable because it pushes back on the most sensational interpretation of agent swarms rather than treating them as magic labor multiplication.
Open at QbitAI Ask Claude ChatGPT

Three frontier models control the same physical camera platform

# Analysis Leiphone 22h

A hands-on test found GPT-5.6 Sol, GPT-6 Astra and Claude Fable 5.1 could move a pan-tilt robot, but differed sharply in coverage, verification and cost.

More
  • Sol covered the full space with a dozen movements, Fable added diagonal corners and reusable tooling, while Astra used seven moves to prove each axis reached its limits.
  • Astra spent more effort verifying feedback, whereas Fable was faster and more optimistic about success; the difference is a policy choice, not simply raw model intelligence.
  • The test exposes a physical-world requirement absent from most coding benchmarks: an action request is not complete until the device reports that it reached a safe state.
  • The cost of certainty becomes visible in hardware tasks because extra motions, checks and retries consume time and can create physical risk.
  • The important comparison is what each model considers evidence of completion: coverage, feedback or a plausible final state.
Open at Leiphone Ask Claude ChatGPT

China adds a 55% tariff on Brazilian beef

# Policy SCMP 11h

China imposed an additional 55% tariff on Brazilian beef after Brazil exhausted its 2026 import quota and failed to secure use of unused Uruguayan quota.

More
  • The measure is a quota-triggered trade action rather than a general diplomatic tariff, with the Ministry of Commerce tying it to the 1.106-million-tonne allowance.
  • It shows how China's agricultural safeguards can rapidly change supplier economics once a national quota is filled.
  • Brazilian exporters' attempt to borrow neighboring Uruguay's unused quota was rejected, underscoring that quota access is politically and legally bounded.
  • The policy may redirect supply toward other markets and raise costs for Chinese buyers, but the supplied report does not quantify the consumer effect.
Open at SCMP Ask Claude ChatGPT

Chinese office agents can do the work but not the accountability

# Analysis Leiphone 22h

A test of Tencent WorkBuddy, Alibaba's office agent and ByteDance's Doubao Work found all could produce spreadsheets, slides and summaries, but quarterly totals differed by 57.7%.

More
  • The three products represent different strategic bets: Tencent emphasizes an extensible platform, Alibaba an organizational entry point and ByteDance enterprise context plus computer control.
  • The dirty spreadsheet contained date errors, refunds, missing departments and anomalies, exposing the gap between generating an artifact and validating business truth.
  • The products are becoming execution layers with local files, permissions and computer access, not simply larger chat windows.
  • The result is especially relevant to managers: automation may remove keystrokes while leaving judgment, checking and liability with the employee.
  • The Chinese market's fragmented messaging, approval and document systems also push domestic agents toward desktop and local-computer control rather than pure SaaS integration.
Open at Leiphone Ask Claude ChatGPT

Chinese scholar reads the Xi-Trump summit as strategic stabilization

# Analysis Ginger River Review 15h

A Chinese policy interview frames the latest summit as progress toward a relationship of constructive strategic stability based on respect, fairness and reciprocity.

More
  • The framing is explicitly political: summit diplomacy is presented as creating a new relationship label, not merely pausing tariff escalation.
  • The interview says both sides discussed reciprocal tariff reductions covering about $30 billion of goods from each country, with most products returning to most-favored-nation rates.
  • It also highlights foreign business and study groups visiting Chinese industrial sites, linking diplomatic stabilization to a campaign to show China's technological future firsthand.
  • The source is a Chinese magazine interview relayed by a China-focused newsletter, so its value lies in elite framing rather than neutral measurement of the summit's outcomes.
Open at Ginger River Review Ask Claude ChatGPT

A free coding agent nearly matches Codex on kernel optimization

# Signal Leiphone 21h

In a controlled test on eight accelerator types, free AgnesCode passed six chips versus Codex's seven and achieved higher speedups on four shared targets.

More
  • The task optimized a DeepSeek V3 fused projection operator, a more demanding test than generating a web page because success required compiling and benchmarking on multiple chips.
  • AgnesCode with Agnes 3.0 Flash took 20 minutes 36 seconds versus 8 minutes 57 seconds for Codex with GPT-5.6 Sol.
  • On shared hardware, AgnesCode reportedly reached 4.81 times speedup versus 3.71 times for Codex on one generic chip, while also passing a Kunlunxin test Codex did not.
  • The result does not show that the free system is generally better; it shows that a bounded task with an automated evaluation loop can let a smaller agent produce professional-grade artifacts.
  • The strategic lesson is that model price matters less when the harness, toolchain and acceptance test can force the job to completion.
Open at Leiphone Ask Claude ChatGPT

Trump says he discussed Chinese political prisoners with Xi

# Policy SCMP 6h

Donald Trump said he discussed political prisoners with Xi Jinping during their summit, while families had urged Washington to press Beijing for releases.

More
  • The statement is a presidential claim, not confirmation that China agreed to release anyone or that a formal mechanism exists.
  • Its timing matters: rights groups and families used the summit before the visit to push an issue that is usually crowded out by trade and security bargaining.
  • If the topic was raised privately, it shows how leader-level diplomacy can create limited space for human-rights cases without changing the broader relationship.
  • The supplied report does not name prisoners or describe a Chinese response, which limits what can be inferred about substance.
Open at SCMP Ask Claude ChatGPT

Xinjiang cotton mechanizes under US sanctions pressure

# Policy SCMP 7h

Xinjiang cotton producers are adopting mechanization and larger-scale farming after US sanctions, raising output while reducing some manual labor but increasing dependence on capital and policy.

More
  • The report contrasts scattered hand-worked plots with mechanized production, showing how sanctions can accelerate substitution and consolidation rather than simply halt exports.
  • A technician's account describes a shift from labor-intensive planting and harvesting to machinery and larger operating units.
  • The human cost is not resolved by mechanization: workers may gain productivity while smallholders lose work or bargaining power.
  • This is a useful case of how Chinese local industries frame sanctions as an incentive for self-reliance, while the external debate focuses on labor rights and coercion.
  • The supplied excerpt is partial, so claims about current output and export destinations require caution.
Open at SCMP Ask Claude ChatGPT

China's AI adoption and enterprise model race accelerate

# Signal GeekPark 5h

A Chinese roundup combines Gemini 4 Argon, Kuaishou's AI reorganization, memory growth and a survey claiming more than 700 million Chinese generative-AI users.

More
  • CNNIC's reported figure would mean generative AI reaches more than half of China's internet users, with question answering the dominant use case and image, video and office work following.
  • The same roundup places Kuaishou's enterprise-AI organization and its Kling leadership change alongside frontier-model news, showing how products and internal operations are being reorganized together.
  • The memory and smartphone claims in the roundup are not independently contextualized, so the strongest standalone signal is the reported user count.
  • The item is a bundle rather than a single event, and should be read as a snapshot of Chinese technology coverage rather than a coherent announcement.
Open at GeekPark Ask Claude ChatGPT

US trade officials accuse China of defying WTO commitments

# Policy SCMP 6h

At a WTO compliance hearing, a senior US trade official accused Beijing of targeting whole industries for global dominance while retaining a non-market system.

More
  • The complaint is about industrial policy as a system, not a single subsidy or trade remedy, so it points toward broader pressure on Chinese sectors.
  • The US framing emphasizes anticompetitive behavior and the limits of WTO rules after 25 years of membership.
  • The practical implication is continued use of trade investigations, allied coordination and sector-specific restrictions rather than a quick settlement.
  • The report supplies the US position but not a Chinese response, so the dispute's evidentiary balance remains incomplete.
Open at SCMP 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 0/20
QbitAI 200 7/10
36Kr 200 3/30
Solidot 200 12/20
Leiphone 200 20/20
GeekPark 200 2/30
IT Home 200 12/60 +42 over cap
SCMP 200 10/50 +8 over cap
Initium 200 3/15
Jiemian 200 8/30 +20 over cap
China Digital Times 200 3/15
The Wire China 200 0/50
Techmeme 200 2/15
Hacker News 200 0/20
V2EX 200 8/50 +31 over cap
Zhidx 200 6/20
InfoQ China 200 8/20 +8 over cap
New Wisdom 200 8/15
TMTPost 200 8/17 +9 over cap