China Desk

China News, Summarized 09 Oct 2026, 15:13 UTC 34 stories 22/23 sources

Today's brief

Beijing halved its own hybrid exports and refused to put the number in writing

Šefčovič got his cap after all

Brussels got the thing it was told it would not get. After two days with commerce minister Wang Wentao in Beijing, EU trade commissioner Maroš Šefčovič said the two sides have a shared understanding to halve Chinese hybrid and plug-in hybrid exports to the bloc — "several million cars" over four years, per Reuters via CNBC. He would not say how it is enforced. "There are different techniques."

Beijing's own text has no number in it. The commerce ministry published a joint statement and a 16-point outcomes list saying only that intensive consultations produced an understanding on hybrid trade "in a manner consistent with WTO rules," via IT Home. A volume cap China spent three months calling WTO-illegal is now a consensus item with no volume in it. Brussels gets to say "half"; Beijing gets to say nothing.

What China took home is in the same list. A green channel for rare-earth and permanent-magnet export licences to Europe, EU facilitation of dual-use licences going the other way, continued talks on cutting tariffs on some goods, and the EV price-undertaking review grinding on. Third meeting March 2027, ministerial video call in January.

I called this wrong yesterday. I wrote that Beijing had decided Europe was bluffing and that I was less sure than it was. Beijing moved instead — against a deficit the EU puts at over €1bn ($1.12bn) a day, and with Porsche announcing 5,000 more job cuts, 9,000 cumulative, a return to combustion engines and 20% price rises, blaming China sales, per Leiphone's wire. EU leaders still have to endorse it on 15–16 October.

Go deeper on this section: ClaudeChatGPT

The State Council outlawed involution by decree

The same day, Beijing told its own factories to stop eating each other. The Party centre and State Council issued opinions on developing "new quality productive forces" — 19 tasks, four principles, science and technology innovation first among five headings, via IT Home. Buried in the emerging-industries section: deepen specialisation within industries and products, build cross-regional cooperation mechanisms, and "avoid involution-style competition."

That word is the tell. Export moderation abroad and anti-price-war discipline at home are the same policy viewed from two ends. The document also hands out the forward list — quantum, biomanufacturing, hydrogen and fusion, brain-computer interfaces, embodied intelligence, 6G — and names intelligent connected NEVs, AI phones and PCs, and humanoid robots as the terminals to push now.

On AI it asks for plumbing, not model cards. Build technology monitoring, risk early warning and emergency response systems; ensure AI is safe, reliable and controllable. Pilot bases and high-value application scenarios, industry by industry. Nothing triggered by what a model can do.

Go deeper on this section: ClaudeChatGPT

ByteDance found DeepSeek answering by token position

Shift the padding, change the answer. ByteDance's Seed team took an FP8 quantisation function out of DeepSeek's own inference code and asked DeepSeek-V4-Flash-Base to complete the last token. Correct answer: 8. Add a decorative docstring of equals signs in front and the model flips to 32 on a four-token cycle, QbitAI reports. Where padding length mod 4 is 0 or 1, the wrong answer averages 71.3% probability; at 2 or 3, the right one averages 91.5%.

At 128K it is worth 40 points. On a haystack of roughly 16,000 key-value pairs, moving the target relative to the compression window swung V4-Flash-Base retrieval accuracy by 40.2 percentage points and V4-Pro-Base by 34.8. Post-training shrinks it — 19.1 for Flash-0731, 14.8 for Pro-0813, 6.1 for V4.1-Flash-0910 — but the ripple survives.

The period is the compression stride. Four tokens for V4, two for V4.1. Seed reproduced the effect from scratch on a 0.6B architecture across several chunked-compression schemes; the full-attention control did not show it, and the cycle tracked the stride whether 4, 6 or 8. Attention heads had quietly specialised by position. Averaged benchmarks hide all of this, which is the point.

Go deeper on this section: ClaudeChatGPT

The poisoning intern just raised thirty million dollars

Place the name first. Tian Keyu is the Peking University doctoral student who, interning on ByteDance's commercialisation tech team in mid-2024, exploited a Hugging Face model-loading hole to slip code into weight files that rewrote colleagues' optimisers and injected random delays — wasting a quarter for some thirty researchers. ByteDance sued for 8m yuan (~$1.1m); the court found breach of the internship contract and ordered 500,000 yuan (~$70,000) plus return of wages.

That December he won the NeurIPS 2024 best paper as first author, for visual autoregressive modelling — next-scale prediction, which kept images two-dimensional instead of flattening them, ran roughly 20 times faster than raster-order autoregression, and showed a clean scaling curve from 300m to 2.3bn parameters. Mainland China's first.

Now ten people, no product, no company name. 5Y Capital, a Moonshot AI backer, and IDG have put close to $30m in at a $200m post-money valuation, Leiphone reports. The pitch: a 200,000-symbol visual vocabulary the model invents for itself, claimed to cut per-second video generation cost by at least an order of magnitude, then roughly 100m hours of video, base model in 2027.

The price is the argument. Leiphone's piece is organised around whether character disqualifies talent. Set that beside the comparables in the same article — Fei-Fei Li's World Labs at a $5.4bn valuation, Yann LeCun's AMI Labs with a $1.03bn seed, former Alibaba model chief Lin Junyang's venture named for the linguistics term pragmatics at $2bn — and $200m for the efficiency thesis looks like cheap seats, character discount included.

Go deeper on this section: ClaudeChatGPT

Washington counted China's missing safety cards

Thirty-one out of 857. SemiAnalysis's report, titled around the claim that Beijing will not pace the frontier, went through every release from nine Chinese developers — Alibaba, ByteDance, Tencent, Baidu, DeepSeek, Moonshot, Z.ai, MiniMax, StepFun — since 2021. Only 3.6% ever carried a published developer safety result; 1.1% had one at launch; 94.9% have none. Where results came, the median lag was 42 days, the longest 349, for DeepSeek-R1.

The report concedes "not found" is not "not tested." Worth holding onto, because the Chinese-language evaluation published the same day was anything but lax. Lenovo's in-house coding harness took first place on SWE-bench-Live Lite at 71% running DeepSeek-V4.1-Flash, and submitted full agent trajectories for the benchmark's audit of answer leakage to earn its verified badge, per QbitAI.

So the gap is in what gets measured, not in rigour. Audited trajectories for capability, a published teardown of a rival's failure mode, no safety card anywhere — and a State Council document that asks for monitoring systems rather than capability-triggered disclosure. SemiAnalysis found the same thing in the binding rules. Both sides are describing one regime; only one of them calls it a gap.

Go deeper on this section: ClaudeChatGPT

Threads we are pulling

  • The hacked banks' toolkit deleted itself. Following yesterday on the South Korean financial breaches: ARTEX's author, posting as Autumn-27, condemned the misuse, said the pentest agent was built for authorised testing, declared it closed-source — and then the whole GitHub repository came down, per Solidot. Reuters has it too. Nothing already cloned is un-published; the engine was Chinese open weights regardless.
  • JAC limit-down twice. Zunjie's statement conceded free upgrades for delivered owners while calling Dongchedi's brake-pedal test "non-standard extreme conditions." ByteDance's auto platform hit back that 100km/h braking is a basic test item and confirmed the viral 4,589N reading was camera-refresh residue against a true peak of 1,612N. JAC closed limit-down again at 22.25 yuan ($3.13), over 10bn yuan (~$1.4bn) gone in two sessions. QbitAI dug out the precedent: Ford recalled 2,627 imported Mustangs in China in 2020 for pedal-bracket fracture, with zero prior complaints on file.
  • Manus got its founder's passport back. Following the $500m round: Solidot reports Xiao Hong's exit ban has been lifted and the post-money valuation is $4bn, led by Boyu and IDG, here.
  • Who absorbs the cost, Taiwan edition. After this week's service-labour pieces: Micron's Taoyuan union won strike authorisation with 1,994 of 2,012 ballots cast — 88.3% of all members — demanding a standing formula of 15% of operating profit shared with staff rather than a bonus, per Solidot. Taichung's union negotiates 22 October and is preparing its own vote.

Go deeper on this section: ClaudeChatGPT

The river · 34

ByteDance researchers find DeepSeek’s long-context accuracy depends on token phase

# QbitAI 8h

ByteDance’s Seed team reports that DeepSeek V4 retrieval accuracy can vary by 40.2 percentage points depending on where information falls inside compressed KV-cache windows.

More
  • The effect repeats every four tokens in V4 and appears tied to the stride used by blockwise KV-cache compression.
  • In a 128K-token retrieval test, moving the same key-value pair changed accuracy dramatically even though the content and context length stayed fixed.
  • Later DeepSeek versions reduce the gap, but V4.1 still shows phase sensitivity, suggesting post-training mitigates rather than eliminates the architectural effect.
  • The mechanism matters for production systems that rely on long-context retrieval: formatting and tokenization can become hidden variables in correctness.
  • This is a concrete example of an efficiency optimization creating a quality artifact at inference time.
  • The work also gives model evaluators a practical robustness test: vary irrelevant prefix length and measure whether decisions change.
Open at QbitAI Ask Claude ChatGPT

Tsinghua-linked robot firm tops a world-action-model benchmark

# QbitAI 11h

Chinese robotics company Astribot says its VPP2 world-action model led RoboDojo in simulation and beat a major baseline in zero-shot real-robot tasks.

More
  • VPP2 reportedly reached 32.26% average success and 39.26 points in RoboDojo, ahead of the cited GPT-6 Astra baseline without extra data or agent-side reinforcement tricks.
  • The model separates video prediction from action learning instead of training them as one undifferentiated objective, aiming to preserve visual generalization while improving control.
  • RoboDojo stresses changed environments, misplaced objects, long-horizon memory and precise manipulation, making it more informative than a fixed demonstration set.
  • On a real dual-arm robot, the system reportedly averaged 58.5% success across ten zero-shot tasks and led in nine, though the report does not provide task-level protocols.
  • The result suggests that world models may need explicit architectural separation between predicting what happens and deciding how to act.
  • It remains a benchmark result, not proof of reliable industrial deployment under uncontrolled contact, lighting and safety conditions.
Open at QbitAI Ask Claude ChatGPT

Manus raises more than $500 million after returning to independence

# GeekPark 14h

Manus says it raised over $500 million at a reported $4 billion valuation after Chinese investors bought it back from Meta, making it a major test of China’s agent financing market.

Also: Solidot, Leiphone

More
  • The round was led by Boyu Capital and IDG, with Tencent, Sequoia China and ZhenFund continuing as investors.
  • The financing follows a politically sensitive reversal after Meta’s reported $2 billion purchase was withdrawn under Chinese regulatory pressure.
  • Manus’s return to independent operation shows that strategic ownership, regulatory jurisdiction and product ambition can matter as much as capital in China’s AI sector.
  • The valuation is reported by people close to the company rather than established through a public filing, so it should be treated as indicative.
  • The company’s next test is whether it can convert financing and recruiting into a durable product moat rather than another cycle of agent hype.
Open at GeekPark Ask Claude ChatGPT

Chinese polar researchers test a Starlink rival with an Arctic-to-Antarctic call

# SCMP 3h

Researchers publicly demonstrated a video call over China’s Guowang satellite constellation, offering the clearest public glimpse yet of a strategic Starlink competitor.

More
  • The test links polar sites rather than ordinary consumer terminals, so it demonstrates constellation capability without proving commercial coverage or capacity.
  • China is accelerating a large low-Earth-orbit network partly because Starlink has shown military and wartime value, turning broadband infrastructure into strategic infrastructure.
  • State media’s emphasis on the call frames Guowang as a national capability milestone, not merely a connectivity experiment.
  • The key unanswered engineering questions are constellation scale, inter-satellite links, terminal availability and spectrum coordination.
  • A successful public test matters less for immediate service revenue than for validating China’s ability to integrate launch, satellites, ground systems and applications.
  • The next meaningful signal will be sustained service and deployment tempo, not another ceremonial demonstration.
Open at SCMP Ask Claude ChatGPT

China’s State Council pushes AI, robots and smart vehicles as new growth engines

# IT Home 1h

A new State Council opinion makes disruptive technology and industrial deployment central to China’s productivity agenda, including AI PCs, smart cars and humanoid robots.

More
  • The document names technology innovation, strategic industries, future industries, industrial modernization and commercialization as five major tasks.
  • It explicitly calls for AI-plus deployment in smart vehicles, AI phones and computers, and humanoid robots while demanding systems for monitoring, warning and emergency response.
  • The policy pairs support for quantum technology, brain-computer interfaces, embodied intelligence and sixth-generation wireless with a warning against cutthroat competition.
  • Its language about avoiding involution signals an attempt to steer local governments away from duplicative subsidy races without abandoning state-led investment.
  • For executives, the important point is demand formation: China is trying to turn emerging hardware and AI categories into coordinated application markets, not merely fund model training.
  • The document still leaves execution to local conditions, so the practical test will be procurement, financing and interoperability rather than the policy list itself.
Open at IT Home Ask Claude ChatGPT

Sharpa builds a tactile humanoid stack from skin to data gloves

# AI QbitAI 10h

Chinese robotics startup Sharpa introduced a humanoid, tactile hands and an exoskeleton glove, tying contact sensing to closed-loop control and training data.

More
  • The D01 humanoid claims 1000 Hz communication, 0.2 mm repeatability and upper-body tactile coverage at 100 Hz, aimed at manipulation rather than stage demonstrations.
  • Its W02 hand reduces active degrees of freedom from 22 to 21 while shrinking the package by about 30%, showing that useful dexterity is an application-specific systems tradeoff.
  • Finger-level visual-tactile sensors and palm electronic skin let the controller respond to contact, slip and force after vision has lost the decisive information.
  • The AE01 glove captures 22 hand-joint signals and feeds force feedback back to the operator, making teleoperation a route to collecting embodied manipulation data.
  • Sharpa’s architecture separates long-horizon task reasoning from high-frequency contact control, a more plausible design than asking one large model to handle every timescale.
  • The commercial test is whether tactile data and control can produce repeatable factory tasks, not whether the hand can perform a memorable demo.
Open at QbitAI Ask Claude ChatGPT

Huawei-backed openJiuwen releases an enterprise AgentOS

# AI QbitAI 12h

Huawei teams and partners released an open enterprise agent operating system with multi-agent orchestration, self-improvement, sandboxing and hardware-aware scheduling.

More
  • The platform treats skills, connectors, plugins, experts and expert groups as reusable assets, aiming to turn business methods into portable components.
  • Its swarm workflow decomposes research, analysis and writing while allowing human intervention at selected checkpoints rather than requiring continuous supervision.
  • The runtime includes multi-tenant isolation, permissions, sandboxes, fault recovery and resource scheduling, addressing the operational layer missing from many agent demos.
  • Huawei is also packaging the stack into private-deployment appliances that it says can be installed in hours, a direct response to Chinese enterprises’ data-locality and control requirements.
  • The open-source strategy could give Huawei influence above its hardware base by making AgentOS an integration point for domestic models and business systems.
  • The report is authorized promotional coverage, so adoption, reliability and ecosystem independence remain unverified.
Open at QbitAI Ask Claude ChatGPT

A premium Chinese car brand responds after brake brackets break in testing

# Policy QbitAI 12h

Seres says its V800 passed standards but will redesign and upgrade delivered vehicles after a test video showed three brake-bracket failures; the tester disputes the company’s framing.

More
  • The company calls the test an extreme nonstandard condition, while the testing outlet says emergency braking at 100 km/h is a basic vehicle evaluation.
  • The disagreement turns on engineering boundaries: a test can exceed a formal procedure without making a structural failure acceptable.
  • The outlet says the actual peak pedal force was 1,612 N, not the widely circulated 4,589 N figure, but repeated failures remain the central unresolved fact.
  • A cited Chinese standard uses 1,500 N bench loading for a particular pedal-strength test, yet that does not directly settle dynamic vehicle behavior.
  • The company’s free upgrade is a meaningful product action, but it is not an explanation of the failure mechanism or proof that the fleet is safe.
  • The episode shows why Chinese premium brands increasingly face independent testing and social-media scrutiny that formal compliance statements alone cannot contain.
Open at QbitAI Ask Claude ChatGPT

Chinese startup raises a record round for diffusion language models

# AI Zhidx 13h

Shenzhen startup DiffuSpace raised several hundred million yuan, roughly $30-60 million, for diffusion language models that generate tokens in parallel rather than left to right.

More
  • The company’s core claim is that global parallel generation could suit edge agents better than autoregressive decoding, especially where latency matters.
  • Its research lineage includes sequence-to-sequence diffusion work, a 7B Dream model and later vision-language-action systems, giving the startup a coherent technical program.
  • The company plans to train and open-source a larger model, while a partner claims a fivefold speedup for edge agents; neither claim is independently established here.
  • Backing from Xiaomi founder Lei Jun’s investment arm, Huawei’s investment unit and Horizon Robotics shows that Chinese capital is treating non-autoregressive language generation as a strategic option.
  • The risk is systems-level: parallel generation may reduce latency but can complicate controllability, iterative refinement and tool-use reliability.
Open at Zhidx Ask Claude ChatGPT

AgentGarten gives agents a code-defined world with neural rendering

# AI QbitAI 13h

MirroS built a real-time training environment where code controls physics and a neural renderer supplies realistic visuals, letting agents act, observe consequences and update strategy.

More
  • The architecture separates deterministic state transitions from image generation: code decides collisions and world state, while the renderer turns geometry and normals into visual observations.
  • That split avoids the two common failure modes of simulators: realistic video with no queryable physics, and precise game worlds that are expensive to make visually rich.
  • The reported system runs above 30 frames per second and lets agents maintain experiment manuals across self-play rounds.
  • In a hide-and-seek recreation, the agents reportedly learned obstacles and ramps within ten rounds, far fewer than the millions of reinforcement-learning games cited for earlier work.
  • The interesting research question is whether language-based reflection and reusable skills can replace some brute-force interaction data without sacrificing physical grounding.
  • The demonstration is a training-environment result, not evidence that agents trained there transfer reliably to real robots.
Open at QbitAI Ask Claude ChatGPT

SCMP says China will sharply limit hybrid-car exports to Europe

# Analysis SCMP 3h

SCMP describes the China-EU understanding as a cut of more than half in hybrid-car exports, contrasting with Beijing’s WTO-compliant trade language.

Also: IT Home

More
  • The export restraint is presented as a temporary mechanism to prevent the electric-vehicle dispute from becoming a broader trade war.
  • The report says the measure would operate over four years, making it a managed adjustment rather than a permanent retreat from European markets.
  • Its framing foregrounds the concession and the political bargain, while Chinese official coverage foregrounds rules, supply-chain stability and reciprocal licensing access.
  • The event shows how Chinese industrial export power is becoming a bargaining instrument in trade diplomacy.
  • Implementation details, including company-level quotas and the exact treatment of plug-in hybrids, remain decisive and are not supplied here.
Open at SCMP Ask Claude ChatGPT

Hong Kong begins a long-delayed reform of sexual-offences law

# Culture Initium 5h

Hong Kong’s consultation drew more than 6,000 responses, with over 90% supporting reform after decades of survivor and advocacy pressure.

More
  • The reform process shows how legal change in Hong Kong can remain stalled for decades before a government consultation creates a formal opening.
  • Survivor groups are pushing not only for updated offences but also for procedures that reduce the burden on victims and improve accountability.
  • The high support rate signals a broad social consensus, but the important question is which protections make it into the final legislation.
  • The report’s Cantonese podcast format also illustrates how Hong Kong media is using local-language reporting for sensitive public-policy debates.
Open at Initium Ask Claude ChatGPT

Lenovo’s TianxiCode claims the top verified SWE-bench-Live result

# AI QbitAI 6h

Lenovo says its TianxiCode agent, paired with a DeepSeek model, solved 71% of verified software-engineering tasks, highlighting the value of agent architecture around a model.

More
  • The benchmark evaluates repository-level issue fixing in reproducible environments, not isolated code completion, so retrieval, planning, tool use and testing all matter.
  • Lenovo describes cross-file retrieval, context pruning, autonomous planning and test-driven self-correction as the system’s main components.
  • The result is attributed to a model-agent combination, making it difficult to isolate Lenovo’s framework from the underlying DeepSeek model.
  • The source is sponsored content, so the claimed ranking and evaluation process should be checked against the official benchmark before treating it as a durable lead.
Open at QbitAI Ask Claude ChatGPT

Xinjiang’s desert seafood boom is sold as a technological miracle

# Analysis The Wire China 2h

China’s Xinjiang fish-farming expansion is presented as proof that technology can overcome geography, while critics point to political control, imported feed and hidden ecological costs.

More
  • The story’s central tension is between a state-backed production narrative and the less visible inputs required to make aquaculture work in a desert environment.
  • Xinjiang’s food projects are also political demonstrations: they turn infrastructure and centralized land control into evidence of development and national integration.
  • Imported feed and water demands complicate the claim that the model is a self-contained technological solution.
  • The case is a useful reminder that Chinese industrial success stories often combine genuine engineering with a carefully managed political frame.
  • What remains unclear from the excerpt is the operation’s profitability and long-term water balance.
Open at The Wire China Ask Claude ChatGPT

A controversial Chinese researcher raises $30 million for a world-model startup

# AI Leiphone 4h

A 10-person lab founded by a disputed former ByteDance intern reportedly raised $30 million at a $200 million valuation to build a symbolic visual language for world models.

More
  • The proposed system uses a 200,000-symbol visual vocabulary to represent and predict video, with the founder claiming an order-of-magnitude reduction in generation cost.
  • The technical bet is to preserve structured spatial and temporal information instead of describing physical scenes only through natural-language tokens.
  • The founder’s earlier work on next-scale visual prediction is a credible research connection, but the new company’s performance claims are not independently demonstrated here.
  • The financing shows that Chinese investors may back technically ambitious founders despite unresolved reputational and governance controversy.
  • The main risk is not only model quality; it is whether a private symbolic representation transfers from video compression to robotics and autonomous driving.
  • The reported funding and valuation come from a profile, so both should be treated as claims rather than verified market facts.
Open at Leiphone Ask Claude ChatGPT

China briefing: EU trade talks, the yuan, gig workers and AI safety

# Analysis Sinocism 17h

A China-focused briefing links EU trade negotiations, the central bank’s yuan stance, gig-worker protections, holiday spending and domestic AI-safety debate.

More
  • The briefing’s value is connective: it places industrial trade, currency policy, labor governance and AI regulation in the same policy environment.
  • Its account of EU talks notes how little Beijing initially disclosed, a recurring feature of Chinese official communication during sensitive negotiations.
  • The gig-worker and holiday-spending items point toward the domestic side of China’s growth problem, where consumption and labor protections matter alongside industrial exports.
  • As a newsletter roundup, it is a map of issues rather than original reporting on each one.
Open at Sinocism Ask Claude ChatGPT

Beijing and Manila trade accusations over a South China Sea evacuation

# Policy SCMP 7h

A medical evacuation near Second Thomas Shoal triggered another China-Philippines dispute, with Manila also reporting a sharp rise in Chinese vessels around contested features.

More
  • The incident shows how even humanitarian or medical activity is now embedded in the operational contest over access to the grounded Philippine vessel.
  • China and the Philippines use different names and legal narratives for the same maritime features, making language itself part of the dispute.
  • The reported near-doubling of Chinese military and coastguard vessels in September suggests a broader pattern beyond the single evacuation.
  • The excerpt does not establish responsibility for the confrontation or whether a communication channel reduced escalation.
Open at SCMP Ask Claude ChatGPT

AWS brings its AgentCore platform to its China regions

# AI Zhidx 8h

Amazon Web Services made its managed agent runtime, identity, browser, sandbox and observability stack available in Beijing and Ningxia, showing how foreign cloud vendors localize enterprise AI infrastructure in China.

More
  • The platform uses microVM isolation, tool gateways, identity management and tracing to package the operational controls enterprises need around agents.
  • Making the service available in Chinese regions matters because data locality, credentials and regulatory boundaries can block otherwise global agent products.
  • AWS claims AgentCore cut infrastructure integration time by 75% in a vendor-sponsored comparison; that result should not be treated as an independent benchmark.
  • The key market question is whether Chinese enterprises will use a foreign control plane for sensitive workflows despite local-cloud competition and compliance constraints.
Open at Zhidx Ask Claude ChatGPT

Bessent may skip Hong Kong finance talks to prepare Trump’s Shenzhen visit

# Policy SCMP 8h

US Treasury Secretary Scott Bessent may prioritize preparations for President Trump’s planned Shenzhen meeting with Xi, keeping the China-US trade truce in focus.

More
  • The report says Bessent is expected to meet Vice-Premier He Lifeng before Trump’s visit, indicating that technical trade negotiations may precede the leaders’ meeting.
  • The immediate objective is reportedly a longer extension of the bilateral trade truce, not a comprehensive settlement.
  • A possible scheduling choice would signal that Washington sees the Shenzhen visit as a higher-value channel than a multilateral finance meeting.
  • Because the account relies on unnamed sources and a future trip, the timing and attendance remain unconfirmed.
Open at SCMP Ask Claude ChatGPT

Wang Yi tells Pacific leaders China’s island ties are not a rivalry

# Policy SCMP 1h

China’s top diplomat framed Pacific engagement as South-South cooperation rather than a contest with Washington, revealing Beijing’s preferred diplomatic narrative.

More
  • The statement was carried by the Communist Party’s International Department, underscoring that party-to-party diplomacy is part of China’s Pacific toolkit.
  • Calling the relationship mutual assistance seeks to distance Beijing’s outreach from the security competition that US and Pacific officials increasingly emphasize.
  • The framing does not remove the strategic reality: aid, infrastructure, diplomatic recognition and security access are being contested in the region.
  • What is missing from the excerpt is the concrete offer made to the island delegation, which would show whether the rhetoric is backed by new commitments.
Open at SCMP Ask Claude ChatGPT

ARTEX disappears from GitHub after abuse claims

# Policy Solidot 6h

ARTEX’s Chinese developer says the security agent was closed and the repository removed after attackers used it against South Korean banks.

Also: Techmeme

More
  • The author says the project was intended for authorized testing and will no longer be updated or supported after the reported misuse.
  • Repository removal changes availability but cannot reliably erase copies, forks or knowledge already distributed.
  • The case is an early example of how Chinese developers are responding to dual-use agent tools under pressure from overseas cyber incidents.
  • Because the account comes from the developer, the attack attribution and scope remain claims rather than independently verified facts.
Open at Solidot Ask Claude ChatGPT

Chinese startup raises about $56 million to build brain-imaging infrastructure

# Analysis GeekPark 10h

Shenzhen startup Kunwei raised 400 million yuan, roughly $56 million, for low-frequency ultrasound that claims to image brain tissue through the skull in real time.

More
  • The company’s engineering bet is to trade higher ultrasound frequency for lower-frequency penetration, then use neural reconstruction to recover spatial detail.
  • The system reportedly processes up to 10 GB of raw acoustic data per second and renders thousands of frames per second, making the bottleneck a real-time signal-processing pipeline rather than a giant language model.
  • Kunwei wants to sell hardware plus a software platform to brain-computer-interface and NeuroAI laboratories, echoing the accelerator-and-software model of a compute infrastructure company.
  • Medical adoption will depend on clinical validation and regulatory claims, not on impressive demonstrations against older ultrasound systems.
  • The report says the company already has tens of millions of yuan in orders, but gives no customer list or independent clinical results.
Open at GeekPark Ask Claude ChatGPT

China’s central bank rejects claims that the yuan is undervalued

# Policy SCMP 6h

The People’s Bank of China says Beijing neither wants nor uses competitive devaluation, directly answering a European criticism during trade negotiations.

More
  • The bilingual policy paper frames exchange-rate stability as compatible with China’s trade growth and rejects the idea that export strength depends on deliberate currency weakness.
  • The timing matters: the statement gives Chinese negotiators a macroeconomic defense while tariff and industrial-policy disputes with Europe intensify.
  • This is a political signal as much as an economic one, because it narrows the language Beijing is willing to accept in any trade bargain.
  • The excerpt does not provide the bank’s empirical estimates of the yuan’s fair value or its response to market-driven depreciation.
Open at SCMP Ask Claude ChatGPT

Xi’s September diplomacy sets up a packed APEC agenda

# Policy SCMP 12h

A South China Morning Post tracker highlights Xi Jinping’s foreign travel and the coming APEC meeting, where he is expected to meet Donald Trump again.

More
  • The item matters mainly as a diplomatic calendar: China will host APEC in November, giving Beijing control over the setting for a high-profile US-China encounter.
  • A Xi-Trump meeting would be the third of the year according to the report, suggesting leader-level contact is being used to manage rivalry rather than resolve it.
  • The excerpt is a tracker rather than an analytical account of concrete agreements, so its value is agenda awareness more than policy detail.
Open at SCMP Ask Claude ChatGPT

AI startups are accelerating acquisitions while Chinese dealmaking remains sparse

# Analysis Zhidx 7h

A Crunchbase review counted 195 AI startup acquisitions this year, while Chinese activity appears limited; the comparison includes Zhipu’s infrastructure purchase and Manus financing.

More
  • The reported 14% increase comes from more acquisitions by roughly the same set of buyers, not a broadening of the buyer population.
  • OpenAI leads the disclosed count, while Chinese companies appear less active in open acquisition data than US and European peers.
  • The comparison is imperfect because the dataset excludes acquisitions by major technology groups, including some Tencent activity.
  • The pattern suggests Chinese AI companies may be relying more on internal development, strategic investment and corporate partnerships than on serial startup purchases.
  • Only 12 of the 195 transactions disclosed prices, so deal counts are more reliable than conclusions about capital committed.
Open at Zhidx Ask Claude ChatGPT

A Chinese diagnostics company’s control fight reaches its board

# Policy TMTPost 2h

Shareholders at IVD company Sansure are split over a board election as a formerly unified control bloc unravels inside a complex partnership structure.

More
  • The controlling shareholder owns 27.53% of the listed company, but voting power depends on who legally controls several partnership vehicles.
  • A 71.88% limited-partner stake does not automatically confer management authority, creating the central money-versus-control mismatch.
  • Five directors rejected an extraordinary meeting because the shareholder request’s seal and representative authority were disputed, not because they opposed succession in principle.
  • The dispute has spilled from partnership litigation into listed-company governance, making board procedure itself a contested asset.
  • The company’s post-pandemic earnings deterioration raises the stakes: control is being fought over while the underlying IVD business remains weak.
  • This is a concrete example of how Chinese listed-company control can be separated from both economic ownership and day-to-day management.
Open at TMTPost Ask Claude ChatGPT

A Chinese think tank urges restraint in the Arctic

# Policy SCMP 2h

A Shanghai-based strategic institute warns Beijing not to confront a new US-Denmark-Greenland security pact or risk exclusion from the emerging Arctic order.

More
  • The advice treats participation and access as more valuable than symbolic opposition, a pragmatic line in a sensitive strategic domain.
  • The report rejects the idea that Washington can make the Arctic an exclusive backyard, while still accepting that the security environment is tightening.
  • Its publication is itself a signal: Chinese policy debate includes arguments for calibrated engagement when direct confrontation would reduce room for maneuver.
  • The excerpt does not show whether this view has official backing beyond the institute’s connection to a university.
Open at SCMP Ask Claude ChatGPT

China’s robotaxi race is shifting from demos to industrial partnerships

# Analysis TMTPost 4h

A Chinese industry analysis argues that robotaxi winners will assemble compute, vehicles, licenses, operations and overseas access rather than own every layer.

More
  • The article describes partnerships linking ride-hailing demand, automaker manufacturing, fleet operations and autonomous-driving stacks.
  • More than 35,000 kilometers of test roads and thousands of permits indicate that Chinese regulators are gradually creating room for commercial deployment.
  • As model-based perception improves and component costs fall, technical demonstrations become a minimum requirement rather than a durable moat.
  • The partnership model spreads capital and regulatory risk but makes integration, data ownership and accountability harder.
  • The analysis is strategic rather than a verified comparison of unit economics, so claims about profitability remain forward-looking.
Open at TMTPost Ask Claude ChatGPT

Tesla suppliers are repurposing China’s auto supply chain for humanoid robots

# Analysis TMTPost 4h

Chinese Tesla suppliers are reportedly moving into humanoid-robot components, using automotive manufacturing, sensing and control expertise as a launchpad.

More
  • The suppliers named are working on joint modules, actuators and precision structures, all areas where automotive production discipline transfers relatively well.
  • Chinese automakers are pursuing several organizational paths at once: internal research, subsidiaries, joint ventures and acquisitions.
  • Factory deployments provide a more valuable training environment than stage demonstrations because robots encounter repetitive but physically variable tasks.
  • The hard problem is moving from one successful operation to reliable, standardized work across many plants and conditions.
  • The article’s market forecasts are promotional and should not be treated as evidence that humanoid robots will soon match automotive scale.
Open at TMTPost Ask Claude ChatGPT

Coal supply disruptions expose China’s thinner import buffer

# Signal TMTPost 4h

Indonesia’s drought-hit exports and India’s low inventories are tightening the global coal market, while China is relying more on Mongolia and other suppliers.

More
  • Indonesia’s low river levels have cut barge loads and trapped coal at mines, making the disruption a physical logistics constraint rather than a policy switch.
  • China’s imports from Mongolia rose 64.6% in the first half, but border throughput limits how much replacement supply can scale.
  • Russian coal exports to China fell despite discounts, showing that logistics, tariffs and route economics matter more than headline price alone.
  • The market signal is a thinner buffer: supply can be redirected, but replacement corridors cannot instantly reproduce Indonesian volume.
  • The figures combine official statistics, vessel tracking and analyst estimates, so exact monthly export totals should be treated as approximate.
Open at TMTPost Ask Claude ChatGPT

Lazada says nearly 10,000 Chinese brands are entering Southeast Asia

# Culture Leiphone 11h

Lazada says almost 10,000 Chinese merchants joined its cross-border program, with Thailand fashion and Vietnam esports showing how localized demand is replacing a generic regional-growth story.

More
  • The platform’s examples point to different go-to-market architectures: lifestyle content and influencers in Thailand, versus teams, streamers and community credibility in Vietnam.
  • Chinese brands are being asked to adapt assortment and messaging to local scenes rather than simply ship domestic bestsellers overseas.
  • The reported growth figures are platform-reported campaign results, so they may not represent durable repeat demand or net profitability.
  • For Chinese companies, Southeast Asia is becoming a portfolio of country-specific channels and subcultures rather than one export market.
Open at Leiphone Ask Claude ChatGPT

A Tang-era calligraphy scroll resurfaces and sells for $22.6 million

# Culture SCMP 13h

A calligraphy work long presumed destroyed resurfaced after 81 years and sold in Hong Kong for $22.6 million, underscoring the market’s appetite for rare Chinese heritage.

More
  • The work is attributed to early Tang calligrapher Yu Shinan and fetched HK$177.45 million, including the buyer’s premium.
  • Its rediscovery gives the sale a provenance story as powerful as the object itself, linking cultural memory, wartime loss and Hong Kong’s auction market.
  • The price is a market signal, not a measure of scholarly consensus or the wider health of Chinese cultural consumption.
Open at SCMP Ask Claude ChatGPT

Chinese biotech executive prosecuted over national drug-bidding collusion

# Policy Jiemian 25m

A subsidiary and controlling shareholder of Chinese biotech company Sano were indicted over bid rigging in a national drug procurement round.

More
  • The alleged scheme involved octreotide injections and illegal proceeds of 10.98 million yuan, roughly $1.5 million.
  • The case shows how China’s centralized procurement system is enforced through both corporate penalties and criminal prosecution, not just price pressure.
Open at Jiemian Ask Claude ChatGPT

Chinese eVTOL startup raises about $14 million as certification begins

# Analysis 36Kr 7h

Qifei Aviation raised nearly 100 million yuan, roughly $14 million, as its five-seat W280 enters Chinese type-certification review.

More
  • The company is using a common platform for passenger and cargo aircraft, hoping to monetize cargo operations before the harder passenger market matures.
  • Its composite-wing design sacrifices some cruise speed for a simpler certification and supply-chain path than tilt-rotor aircraft.
  • The W280 targets a 30 to 150 km urban and intercity market, with a claimed 300 km range and 3.18-ton maximum takeoff weight.
  • The financing is backed by provincial, municipal and district state capital, showing how local governments are underwriting strategic aviation experiments.
  • Type-certification acceptance is an important milestone, but it is not approval for commercial passenger operations.
Open at 36Kr 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 0/20
Ginger River Review 200 0/20
Sinocism 200 1/20
QbitAI 200 10/10
36Kr 200 5/30
Solidot 200 13/20
Leiphone 200 11/20
GeekPark 200 3/30
IT Home 200 12/60 +42 over cap
SCMP 200 10/50 +8 over cap
Initium 200 4/15
Jiemian 200 8/30 +15 over cap
China Digital Times 200 0/15
The Wire China 200 1/50
Techmeme 200 3/15
Hacker News 200 0/20
V2EX 200 8/40 +31 over cap
Zhidx 200 10/20
InfoQ China 200 8/20 +10 over cap
New Wisdom 200 8/15
TMTPost 200 8/16 +8 over cap