China Desk — 14 Aug 2026, 09:15 UTC

GLM-5.3 targets coding, agents and security

# QbitAI 27m

Chinese testing says GLM-5.3 is the strongest open model on several benchmarks, while its red-team program found 2,404 vulnerabilities across 220 software projects.

The report says coding accuracy continues to improve with larger reasoning budgets, suggesting that inference-time compute remains a central lever for the model.
A CyberGym white-box code-review score of 84.5 percent is reported, up from 77.2 percent for GLM-5.2, though the comparisons are outlet-reported rather than independently audited here.
Zhipu says it conducted intensive red-team work with universities, companies and laboratories before release and found 1,088 medium- or high-risk flaws among the screened results.
The reported flaw set spans kernels, operating systems, browser engines, open-source components, applications and protocols, making the exercise more like ecosystem security testing than a model-only audit.
A demonstration built a 3D scene and a small authenticated game, but demos do not establish reliable production performance.
The strategic shift is notable: Chinese model vendors are marketing security evaluation and long-horizon tool use alongside leaderboard scores.

Open at QbitAI Ask Claude ChatGPT

China moves to tax offshore trusts used by wealthy residents

# Pekingnology 6h

New rules tax offshore trusts at transfers, income and distributions; a Tsinghua economist presents the move as both a fairness measure and a case for lowering top labor-tax rates.

The Ministry of Finance and State Taxation Administration rules close timing and structuring routes that allowed Chinese tax residents to defer or avoid personal income tax through offshore trusts.
The article says offshore trust assets are estimated at about 2.3 trillion yuan, roughly $324 billion, though the estimate is attributed to industry sources.
Professor Li Daokui argues ordinary people are largely unaffected because they do not hold offshore trusts, while wealthy residents face a more direct compliance burden.
He also argues that taxing capital income at up to 20 percent while labor income can reach 45 percent is inequitable, turning a crackdown into an argument for broader tax reform.
The policy arrives amid fiscal pressure and a stronger state interest in revenue collection, but the source does not quantify expected receipts.
The broader signal is that offshore wealth is becoming less politically tolerable and less operationally safe for Chinese tax residents.

Open at Pekingnology Ask Claude ChatGPT

DeepSeek V4 Pro's post-training jump makes agents usable

# GeekPark 1d

Chinese hands-on testing finds DeepSeek V4 Pro's architecture unchanged but its agent benchmarks dramatically improved after post-training, narrowing the gap with expensive closed models.

The model remains a 1.6-trillion-parameter MoE with 49 billion active parameters, one-million-token context and 384,000-token maximum output, according to the report.
Reported scores rose sharply on software engineering, CyberGym, Terminal Bench and a hard agent benchmark, indicating that the release targeted tool use and long code chains rather than base architecture.
A real coding test produced complete interactive simulations, but a thinking mode used most of its output budget for reasoning and truncated the final code; disabling thinking fixed that task.
The model is not presented as universally best: it trails leading closed models on some knowledge and repository tests while offering a much lower price.
The commercial significance is the combination of open access, low inference cost and newly credible long-horizon execution.
The release also validates Harness as part of the model strategy: DeepSeek is optimizing the model and the runtime together.

Open at GeekPark Ask Claude ChatGPT

Zhu Rongji, architect of China's economic reform, dies at 97

# SCMP 2h

Zhu Rongji's death has prompted Chinese and international retrospectives: he centralized fiscal power and drove market reform, while leaving memories of industrial upheaval and social cost.

Also: Initium, The Wire China

Zhu's legacy sits at the intersection of reform and centralization: he pushed market restructuring while strengthening the center's control over revenue and macroeconomic policy.
The Chinese account emphasizes his anti-corruption image, forceful personality and role in developing Shanghai and Pudong before he entered the national leadership.
The reform program helped create a more governable fiscal state, but the source also links it to harsh memories among industrial workers and affected families.
His career is useful context for understanding why today's Chinese economic governance combines market mechanisms with strong administrative command.
The competing memory of efficient reformer and architect of social dislocation explains why his death carries more political meaning than a routine former-leader obituary.
The long-term question is whether the state can preserve central coordination without repeating the social costs associated with his restructuring era.

Podcast revisits Zhu Rongji's reform legacy Initium
Remembering China's reform-era premier Zhu Rongji The Wire China

Open at SCMP Ask Claude ChatGPT

Apple and Alibaba train a China-specific language model

# Solidot 1h

Chinese coverage frames Apple's China model as a route through local regulation and competition, while US reporting focuses on Apple becoming the first foreign firm with a proprietary model there.

Also: Techmeme

The reported model is being trained with Alibaba's support for Apple devices sold in China, where Apple Intelligence still requires local compliance and registration.
The Chinese framing emphasizes control over the user experience as Huawei and other domestic rivals gain ground with AI-enabled phones.
The US framing makes the partnership a sovereignty story: Apple needs a Chinese technology partner and locally acceptable model behavior to ship its own branded AI.
This is not simply localization of an interface; it indicates that model choice, training support and regulatory access are becoming part of a foreign platform's country architecture.
The arrangement also shows how China's regulatory boundary can turn a global product into a partially separate technical stack.
The most important unknown is how much of the model, data pipeline and safety layer Apple controls versus Alibaba and other local partners.

Reuters: Apple trained a China-specific model with Alibaba Techmeme

Open at Solidot Ask Claude ChatGPT

China names 75 apps for illegal personal-data practices

# IT Home 37m

A joint Chinese enforcement campaign identified 75 apps and mini-programs for privacy-policy, consent, third-party sharing and account-deletion violations.

The inspection covers mobile apps and mini-programs, with failures including inadequate first-run disclosure, incomplete descriptions of embedded third-party code and missing consent.
The listed problems also include collecting data before permission, failing to offer effective deletion or account cancellation, and sharing personal information without required disclosure.
The action is part of a coordinated campaign involving cyberspace, industry, police and technical-response authorities, showing that app privacy enforcement is becoming routine multi-agency work.
For product teams, the risk is not limited to a privacy statement: permissions, SDK inventories, data recipients and exit mechanisms all become enforcement surfaces.
The breadth of the list suggests a compliance regime focused on operational implementation rather than only major data breaches.
The strategic implication is that Chinese consumer software must treat privacy controls as a launch and distribution requirement, not a later legal layer.

Open at IT Home Ask Claude ChatGPT

DeepSeek Harness is a developer-first, self-evolving agent shell

# AI QbitAI 17h

Chinese hands-on testing praises Harness's pluggable runtime, event-level trajectory view and long-task behavior, while noting a less polished interface than Codex.

Also: GeekPark, Leiphone, Hacker News

The framework exposes models, tools, policies, storage and context management as interchangeable components, enabling scenario-specific runtimes rather than a fixed assistant.
Its trajectory view records raw events and token spending, making agent behavior more inspectable than a polished chat transcript.
The tester found that Harness often kept working through long coding tasks with fewer half-finished outputs, but this is an individual comparison rather than a controlled benchmark.
The current product is strongly oriented toward developers: it runs through a browser UI, lacks a native desktop app and has visible interaction gaps.
A built-in token dashboard and cache-hit reporting point toward a future in which agent orchestration and cost governance are part of the same runtime.
The release matters because DeepSeek is trying to own the control plane around its models, not merely distribute model weights.

Open at QbitAI Ask Claude ChatGPT

What DeepSeek's open-model strategy is really trying to do

# Analysis ChinaTalk 22h

ChinaTalk's analysis portrays DeepSeek as a Hangzhou lab pursuing automated learning and generalized intelligence through open models, with a business model built around enterprise services.

The analysis says Liang Wenfeng treats advanced machine intelligence as the central problem and is willing to trade near-term commercial optimization for that research direction.
Open weights are presented not as a lack of business model but as a way to spread adoption while monetizing enterprise deployment, support and related services.
The essay places DeepSeek inside China's political economy: hardware constraints are real, but the lab's horizon is defined by learning efficiency and long-term capability rather than current chip access alone.
Its comparison with Western AI leaders is useful but interpretive, based partly on leaked meeting minutes whose completeness and authenticity are not established in the excerpt.
The strategic difference is organizational as much as technical: a concentrated, founder-directed research culture can pursue a theory of action that would be difficult for a conventional product company.

Open at ChinaTalk Ask Claude ChatGPT

Baidu spins its AI office agent into a standalone product

# AI QbitAI 39m

Baidu says its Kuku AI office agent has more than 25 million monthly active office users and is moving into dedicated desktop, web, mini-program and enterprise products.

The product combines a large document corpus with private cloud files, persistent memory, skills and sub-agents for Word, spreadsheets, presentations, code and media.
Its architecture is described as MoE-backed, tool-driven and multimodal, with enterprise governance, identity controls and audit trails added to the enterprise edition.
The company is targeting finance first, where it offers research, valuation, meeting and spreadsheet skills rather than only conversational assistance.
The 25 million user figure and national evaluation rankings are company-provided, and the article is authorized promotional material, so adoption quality and retention are not independently established.
The strategic bet is that the durable office product is an agent connected to organizational knowledge and workflows, not a generic chat window.

Open at QbitAI Ask Claude ChatGPT

Claude may have helped solve every sub-2000 Hadamard gap

# Signal QbitAI 21h

Chinese coverage says a team including Claude constructed 12 previously missing Hadamard matrices below order 2,000, including the long-standing order-668 case.

The matrices use only plus and minus one entries with pairwise orthogonal rows, making the result a discrete construction problem rather than a language benchmark.
The report says exact integer checks reproduced all 12 constructions, but the research community has not yet established whether the method is general or a search-specific breakthrough.
EpochAI provisionally labels the 668 problem AI-solved while reserving judgment pending a technical paper and independent understanding of the construction.
The human-AI division of labor remains unclear: the team included three human researchers and Claude, and the source says the model's discovery process is not yet known.
If the construction generalizes beyond the 12 cases, it would be more important than a one-off answer because it could advance combinatorial design methods.

Open at QbitAI Ask Claude ChatGPT

Why Chinese users say online speech is swallowed by filters

# Policy China Digital Times 8h

A translated Chinese post argues that blunt platform keyword filtering blocks harmless speech and forces users into coded language, while framing the problem as platform overcompliance.

The post distinguishes national content rules from the much larger set of terms platforms may block to avoid enforcement risk, placing user experience inside a layered responsibility system.
Users reportedly resort to homophones, initials and coded expressions because ordinary discussion can trigger automated filters without a clear public directory of prohibited words.
Blaming platform design rather than only state policy is strategically significant: it identifies a narrower space where companies could reduce needless censorship without openly challenging the legal system.
The account remains a critique from one online post, not evidence that every platform applies the same rules or that national authorities endorse each individual block.

Open at China Digital Times Ask Claude ChatGPT

Robot training demand is rising five to tenfold

# AI QbitAI 1h

A Chinese robotics team says its compute demand grew five to ten times in two years, illustrating why embodied-AI development is becoming a cloud infrastructure problem.

The workflow combines simulation, data generation, model training, multi-user collaboration and real-robot validation, unlike a conventional single-model inference service.
The reported bottlenecks are systems-level: CUDA, ROS2, simulators, drivers and dependencies must be made reproducible across teams and hardware.
Cloud workstations let teams allocate GPUs elastically instead of buying fixed local machines that may sit idle after a project.
The provider's claim of 80 percent deployment-efficiency gains is promotional, but the underlying problem is structural: engineering friction grows with team size and scenario complexity.
Embodied AI therefore needs a shared environment layer spanning compute, 3D interaction, data permissions and deployment, not merely more accelerator capacity.

Open at QbitAI Ask Claude ChatGPT

DeepSeek opens Harness while Chinese chip stocks surge

# AI GeekPark 8h

A Chinese tech roundup leads with DeepSeek's MIT-licensed Harness preview, but also bundles unrelated claims about OpenAI revenue, Anthropic valuation and Changxin's market value.

The Harness section describes an all-plugin agent runtime in which models, tools, UI, sandbox and scheduling can be replaced without editing core source code.
The roundup's other market and foreign-company claims are secondary items, so their inclusion should not be read as a single verified thesis about China's AI industry.
Its useful signal is that Chinese coverage presents open agent infrastructure and open models as part of a wider contest over domestic compute and capital markets.

Open at GeekPark Ask Claude ChatGPT

Acrab raises $130 million for agent-focused edge AI chips

# AI QbitAI 20h

Singapore-based Acrab raised $130 million after bringing its first edge AI chip into production, betting on local inference for persistent, privacy-sensitive agents.

The GΞLIX 1 platform combines CPU and NPU resources, unified memory and prefill optimization rather than selling peak TOPS alone.
The company reports 700 TOPS and 1,416 tokens per second on a long-context Gemma test, but those figures are self-reported.
The architecture targets a split workload: heavy reasoning remains in the cloud while high-frequency, latency-sensitive and private tasks stay on the device.
Agent workloads create bursty CPU-GPU transitions and repeated tool calls, making memory bandwidth, scheduling and software integration as important as accelerator throughput.
The funding and product timing suggest investors see edge infrastructure as the next layer after data-center AI, although customer-scale deployment is not established.

Open at QbitAI Ask Claude ChatGPT

Embodied-AI data startup raises millions within 40 days

# AI QbitAI 21h

Chinese physical-AI infrastructure startup Yuandian raised tens of millions of yuan and is building a data network for real-world robot interactions, where high-quality data remains scarce.

The company claims the domestic embodied-AI sector raised 93.5 billion yuan, roughly $13.2 billion, in the first half of 2026, while usable high-quality physical interaction data remains only about 500,000 hours.
Unlike web text, robot data must align vision, touch, joint trajectories, object physics and temporal context through real-world collection.
Its MatrixOS platform aims to manage collection, processing and distribution while serving robot makers, world-model companies and solution providers.
The reported million-scale customer order and university partnerships indicate early demand, but the company's claims about global scale and data scarcity are not independently substantiated here.
The strategic opportunity is to become neutral infrastructure for a fragmented data supply chain, provided its schemas and quality metrics win adoption.

Open at QbitAI Ask Claude ChatGPT

Open-source stack makes Qwen3.8 usable across nine AI chips

# AI Leiphone 1d

FlagOS says it adapted Alibaba's 2.4-trillion-parameter Qwen3.8 to nine Chinese and foreign accelerators within a day, using quantization and a common software layer.

The model has 2.4 trillion total parameters and 95 billion active parameters, with native 262,144-token context expandable to about one million tokens.
FlagOS says it supports BF16, FP8 and INT8 paths across chips from Nvidia, Huawei, Moore Threads, Hygon and other vendors, with precision aligned against a CUDA FP8 reference.
The engineering problem is not only compiler compatibility: a model six times larger than the prior 397-billion-parameter version stresses memory capacity, bandwidth, interconnect and parallel scheduling.
The project's Day0 model is intended to keep existing Chinese accelerator clusters useful as new models arrive, improving utilization and reducing the risk that hardware becomes obsolete before its planned life ends.
The reported benchmarks and 24-hour deployment claims come from the open-source project and its partners, so independent reproducibility remains important.
The strategic implication is that a shared software stack can reduce vendor lock-in and make heterogeneous domestic compute commercially viable.

Open at Leiphone Ask Claude ChatGPT

What China's AI talent market is actually paying for

# Culture Ginger River Review 15m

A Chinese podcast discussion suggests million-yuan offers for new AI PhDs are exceptional; firms are paying for scarce systems, data and deployment skills as hiring spreads beyond labs.

The discussion distinguishes headline compensation from the broader market, where companies are competing for specific capabilities rather than every AI worker.
The recruiting firm's client base spans chips, infrastructure, foundation models, robotics and agents, indicating that the shortage is distributed across the stack.
The more consequential shift is organizational: AI expertise is moving into finance, manufacturing and other sectors where integration and domain knowledge matter as much as model research.
The source is a curated podcast excerpt rather than a labor-market survey, so its numbers describe recruiter experience rather than a representative wage distribution.

Open at Ginger River Review Ask Claude ChatGPT

China is urged not to make its AI body an anti-US bloc

# Policy SCMP 46m

Analysts cited by Chinese coverage warn that the new World Artificial Intelligence Cooperation Organisation must avoid appearing naturally opposed to the US and its allies.

The argument is that prospective members may avoid an institution perceived as a geopolitical counter-alliance rather than a technical cooperation forum.
The comparison with the US-backed Pax Silica partnership frames AI governance as an emerging standards and dependency contest.
China's challenge is to use institutional influence without making membership a loyalty test.

Open at SCMP Ask Claude ChatGPT

SHEIN's IPO story is about redemption rights, not cash

# Analysis 36Kr 4h

Chinese analysis argues that SHEIN is pursuing an IPO despite $14.8 billion in cash because $17.3 billion of preferred stock carries reactivated redemption rights.

The company generated about $6 billion in operating cash over 2023 to 2025, more than its roughly $3.8 billion of equity financing over the prior eleven years.
The analysis says preferred-stock redemption economics changed in March, creating pressure if the IPO is withdrawn, rejected or delayed beyond the specified window.
SHEIN's cash conversion model relies heavily on customer prepayments and fast inventory turnover, which can fund growth without conventional financing.
Its valuation history shows a sharp reset: the reported D-plus round priced the company at $64 billion, 35 percent below the prior round despite continued operating growth.
Revenue growth has slowed sharply, with US revenue reportedly declining in 2025 and again in the first quarter of 2026.
The strategic question is whether public-market access can refinance a complex capital structure while investors reassess the durability of the fast-fashion growth engine.

Open at 36Kr Ask Claude ChatGPT

RISC-V startup heads for Hong Kong IPO after losing 5.2 billion yuan

# Analysis 36Kr 4h

A Chinese chip startup founded by former BOE chairman Wang Dongsheng is seeking an IPO after three years of losses and heavy state-linked funding.

The company reportedly raised 9.28 billion yuan, about $1.3 billion, across four rounds while accumulating 5.2 billion yuan, roughly $730 million, in losses from 2023 through the first quarter of 2026.
Its stated strategic bet is replacing x86 and Arm in intelligent-terminal control chips with RISC-V, but its all-market share remains small and the domestic-label ranking flatters its position.
The reported 2025 gross profit of 453 million yuan did not cover 604 million yuan of administrative expense, before 1.04 billion yuan of research spending.
Cash fell to 188 million yuan at the end of 2025, while later liquidity was supported by borrowing and liquidating financial products rather than operating cash generation.
One customer reportedly supplied 82.1 percent of 2023 revenue, declining to 39.3 percent in the first quarter of 2026, leaving concentration risk even as it improves.
The IPO is therefore a financing and industrial-policy test: strategic architecture ambitions still need commercial margins, diversified customers and a survivable cash cycle.

Open at 36Kr Ask Claude ChatGPT

Kimi's IPO case depends on turning model quality into capital

# Analysis 36Kr 5h

Chinese analysis says Moonshot AI needs a new IPO narrative: Kimi K3 offers frontier performance at lower cost, but compute shortages and fast rivals threaten the window.

The article treats a public listing as more than fundraising: it would create a recurring source of capital for compute, model iteration and future acquisitions.
Kimi K3's reported benchmark position gives Moonshot a technical story, but DeepSeek's newer model reportedly offers a much lower per-task cost and weakens that advantage quickly.
A temporary pause on new consumer subscriptions after K3's launch exposed a basic constraint: strong demand is not useful if the serving cluster cannot absorb it.
The company therefore faces a race between technical visibility and infrastructure capacity, with public-market timing serving as a way to convert temporary model leadership into financing.
The analysis also argues that the leading labs share similar Transformer-derived techniques, making sustained differentiation harder than a single leaderboard rank suggests.

Open at 36Kr Ask Claude ChatGPT

Tencent's AI spending is becoming a storage bet

# Signal 36Kr 23h

Tencent's second-quarter capital spending rose 176 percent year on year to 52.8 billion yuan as it bought compute and memory for larger models, pushing free cash flow negative.

Second-quarter revenue was 204.8 billion yuan, about $28.8 billion, while capital expenditure was roughly $7.4 billion and research spending rose 35 percent.
The analysis argues that AI infrastructure requires not just GPUs but high-bandwidth memory and storage, making Tencent a demand source for a supply-constrained memory market.
Tencent's Hy3 and future larger models increase pressure on memory bandwidth, capacity and latency, while storage expansion typically takes years to bring online.
The spending reflects a strategic use of mature games and advertising as cash engines for a new AI growth curve, but investors reacted negatively to near-term margin and cash-flow pressure.
The article is investment analysis rather than a company forecast, and its claim that Tencent will sustain a storage bull market should be treated as interpretation.
The durable signal is that Chinese internet firms are now committing capital at infrastructure scale, not merely adding AI features to existing apps.

Open at 36Kr Ask Claude ChatGPT

Chinese STEM students turn to Singapore as Western options narrow

# Culture SCMP 11h

Chinese students are increasingly considering Singapore as US, UK, Canadian and Australian study routes become less predictable amid technology rivalry and immigration restrictions.

The shift is driven by both push and pull factors: tighter Western access and Singapore's proximity, regional role and relatively stable policy environment.
For China, this is a talent-mobility change rather than just an education trend; universities and employers in Singapore can capture students who once fed Western research systems.
The source presents a directional pattern but does not quantify applications, enrollment or the share of students changing plans.

Open at SCMP Ask Claude ChatGPT

Chinese electric vehicles reach Brazil through a local operating network

# Policy Leiphone 23h

Didi's Brazilian platform 99 and rental company Movida plan discounted EV leasing for 50,000 drivers, linking Chinese vehicles to an established local mobility ecosystem.

The plan is designed to reduce driver rental costs by about 25 percent, addressing the upfront barrier that prevents lower operating costs from translating into EV adoption.
The reported Brazilian sustainable-mobility alliance includes 31 companies across vehicles, charging, finance and rental, with more than 410 million reais invested since 2022.
This is a different export model from simply shipping cars: a Chinese-linked platform supplies demand, drivers, financing relationships and operational data around the vehicles.
The model resembles China's own path, where ride-hailing fleets and charging networks helped turn EVs into high-mileage commercial tools before mass consumer adoption.
The strategic implication is that Chinese companies can export an operating system for industrial adoption, not just a product.

Open at Leiphone Ask Claude ChatGPT

Microsoft is shrinking in China but keeping strategic customers

# Policy Solidot 5h

Chinese coverage says Microsoft closed at least 15 China offices and joint ventures, yet remains to serve firms such as ByteDance and retain engineering talent.

The article links the retrenchment to China's preference for domestic software, US export controls and slower market growth for Microsoft's cloud and AI businesses.
Microsoft's remaining value proposition is two-sided: Chinese companies need Western technology to manage global operations, while Microsoft wants access to China's engineering workforce.
The result is not a clean withdrawal but a narrower China strategy shaped by regulation, geopolitics and customer-specific profitability.

Open at Solidot Ask Claude ChatGPT

China warns against online myths that deny Western history

# Culture SCMP 8h

A Zhejiang propaganda account warned against pseudo-history theories claiming Aristotle and ancient Greece never existed, showing official concern about nationalist conspiracy communities.

The intervention targets online narratives that say the West stole major civilizations and that canonical European history is fabricated.
The state is not simply defending a historical curriculum; it is trying to prevent uncontrolled conspiracy narratives from competing with its own account of national history.
That makes the episode a useful signal of the limits of online nationalism: mobilizing anti-Western sentiment is acceptable only when it remains politically legible and controllable.

Open at SCMP Ask Claude ChatGPT

China tests a $10.8 billion canal link to Southeast Asia

# Policy SCMP 5h

A Chinese infrastructure project completed its first full transit test along a 134.2-kilometer canal, highlighting Beijing's inland logistics and regional-connectivity strategy.

The source describes the project as a major canal link from inland China toward Southeast Asia, with a reported cost of $10.8 billion.
A full-route test is an operational milestone, but the excerpt gives too little information to assess commercial traffic, environmental costs or financing.

Open at SCMP Ask Claude ChatGPT

China rejects New Zealand's latest spy assessment

# Policy SCMP 2h

Beijing's embassy in Wellington called New Zealand's threat report foreign interference and Cold War thinking after the agency named China its only espionage threat operating at scale.

The dispute shows how security assessments now become diplomatic incidents, with China contesting not only individual allegations but the legitimacy of the threat framework.
New Zealand's report reportedly singled out China at a time when the bilateral relationship is already strained by espionage and geopolitical concerns.
The disagreement leaves little room for technical confidence-building: Wellington frames the issue as state activity, while Beijing frames the assessment itself as politicized.

Open at SCMP Ask Claude ChatGPT

Old AMD processors expose protected memory to privileged attackers

# Signal IT Home 44m

A researcher demonstrated a memory-address remapping attack against AMD Family 16h processors that can reach protected firmware and security regions; AMD says no patch is planned.

The technique flips an unlocked memory-controller configuration bit to create an alias address mapping, allowing access to otherwise protected DRAM.
Potential targets include the platform security processor, system-management memory, suspended CPU state and stored microcode, according to the report.
The affected chips include older Jaguar and Puma-era APUs; modern Zen processors are not affected by the reported issue.
Exploitation requires root or administrator privileges, so this is a post-compromise escalation and insider-threat problem rather than an initial-entry vulnerability.
AMD confirmed the issue in security bulletin AMD-SB-7068 and said the discontinued families will not receive a fix.
The architectural lesson is that hardware isolation can fail below the operating system when configuration registers governing physical address translation are not locked.

Open at IT Home Ask Claude ChatGPT

AI opens a new chapter for ancient Chinese literature

# Culture SCMP 5h

Chinese coverage describes AI tools for deciphering, punctuating, comparing and preserving ancient texts, changing a labor-intensive scholarly workflow.

The technology is being applied to fragile manuscripts and texts written without modern punctuation, where specialists traditionally spend years on transcription and comparison.
The important cultural issue is not just automation: machine-assisted recovery can expand access to sources while raising questions about editorial authority and errors in interpretation.

Open at SCMP Ask Claude ChatGPT

Guangdong launches loans backed by AI token consumption

# Policy IT Home 30m

A Guangdong bank is offering credit tied to contracted AI token usage, with preliminary approvals of 28 million yuan, turning model consumption into a financing metric.

The reported pilot can combine unsecured credit, receivables pledges and order financing, with loan limits linked to a firm's contracted token consumption.
At 28 million yuan, the preliminary lending totals roughly $3.9 million under the requested conversion rule.
The policy experiment treats model access and compute consumption as operating inputs that can support credit underwriting, a step toward financializing AI infrastructure demand.
The source does not say how default risk, volatile usage or provider concentration will be handled, so the product's durability remains unclear.

Open at IT Home Ask Claude ChatGPT

Chinese tech roundup: Changxin overtakes Tencent in market value

# Culture 36Kr 9h

A Chinese roundup highlights domestic memory maker Changxin surpassing Tencent in market value, tiny retail subscription losses for Unitree and a bank-based marriage office.

Changxin's reported market value reached about 3.54 trillion yuan, roughly $499 billion, after a sharp post-listing rally, reflecting speculative enthusiasm for China's AI and memory supply chain.
The Unitree IPO detail is a small demand signal: 8,734 online shares were abandoned, only about 0.03 percent of the non-strategic offering.
A marriage-registration office inside a Tianjin bank shows how Chinese institutions blend civic services with financial marketing, including wedding certificates and household financial planning.
The roundup also repeats the DeepSeek Harness preview, but that separate event is better treated under the dedicated Harness cluster.
The market figures are volatile and the roundup provides little fundamental explanation for Changxin's valuation.

Open at 36Kr Ask Claude ChatGPT

Chinese regulator warns biotech firm over $19 million fund diversion

# Policy Jiemian 33m

Xinjiang's securities regulator ordered Chen Guang Bio to rectify a 136 million yuan non-operating fund occupation and issued a warning that enters the company's integrity record.

The funds were borrowed by consolidated subsidiaries in March and April 2025 and repaid with interest in May, according to the filing.
The amount is roughly $19 million, and the regulator's action includes a written correction report and a formal warning rather than a reported operating shutdown.
The case shows how Chinese listed-company oversight reaches through consolidated subsidiaries when related-party cash movements are not properly controlled.

Open at Jiemian Ask Claude ChatGPT

US accuses nine Latin American states of enabling tariff evasion

# Policy SCMP 9h

Chinese coverage reports that Washington added Mexico, Panama and Colombia to a list of countries at elevated risk of routing Chinese goods around US tariffs.

The policy targets transshipment through third countries, broadening the trade dispute from Chinese exporters to logistics and customs systems across Latin America.
The accusation also puts pressure on governments that recently moved against Chinese commercial interests, making tariff enforcement part of a wider geopolitical bargain.
The source reports the US position but does not include named shipments or the affected product categories.

Open at SCMP Ask Claude ChatGPT

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	10/10
36Kr	200	9/30
Solidot	200	13/20
Leiphone	200	13/20
GeekPark	200	3/30
IT Home	200	12/60	+40 over cap
SCMP	200	10/50	+8 over cap
Initium	200	4/15
Jiemian	200	8/30	+16 over cap
China Digital Times	200	1/15
The Wire China	200	1/50
Techmeme	200	4/15
Hacker News	200	1/20
V2EX	200	8/50	+40 over cap

DeepSeek spent two years being the floor under everyone's prices. On Monday it stops.

What Beijing is signalling

The AI race

The vibe

Read differently

Threads we are pulling