China Desk

China News, Summarized 25 Aug 2026 34 stories archived day

Today's brief

The anti-graft campaign is not winding down — it is being handed foreign jurisdiction

A 47-article draft gives the corruption hunt a passport

The National Supervisory Commission wrote itself a statute. China's top legislature took a first reading Tuesday of a draft Anti-Cross-Border Corruption Law — six chapters, 47 articles, text not yet public — at the 24th session of the 14th National People's Congress Standing Committee, which runs to Friday.

Xinhua's framing is about recovery. The law is meant to enrich the legal toolbox for fugitive repatriation, asset recovery and foreign-related cases, closing jurisdictional gaps on offshore offences. 963 fugitives were repatriated in 2025, per the discipline commission's own work report.

But the draft also sets out obligations for companies. Integrity and compliance duties, and the liabilities for breaching them. That is the part aimed at Chinese firms operating abroad rather than at absconded cadres — a compliance regime with Beijing as the enforcing jurisdiction, on top of whichever local one applies.

And the campaign being codified is speeding up, not settling. SCMP's data dive counts 65 probes into civilian officials of vice-ministerial grade or above last year against 19 in 2013, Xi's first year. The Central Commission for Discipline Inspection, the Party's internal anti-graft body, has opened 46 investigations so far this year against 39 by this point last year.

The personnel move came first. The commission held an out-of-cycle Sixth Plenum on Saturday — these normally convene in January — naming Zhao Shiyong and Zhang Zhong deputy secretaries in place of Xiao Pei and Fu Kui, both associates of Wang Qishan, who ran the commission through Xi's first term. Sinocism reads it as preparation for the coming Fifth Plenum. Hard to disagree.

Go deeper on this section: ClaudeChatGPT

Washington prices one H-1B at $103,265

The fee is back, by rulemaking this time. The Department of Homeland Security published a proposed rule establishing a $103,265 charge on every cap-subject H-1B petition, payable at filing, on top of every existing fee — more than a hundred times the base petition fee, and derived by dividing about $8.78bn of claimed annual government cost by 85,000 petitions. Universities and nonprofit research bodies are exempt. Comments close 24 September.

A federal judge struck down the earlier $100,000 version in June. Solidot's summary for Chinese readers pairs it with the other plan: revoking up to 200,000 business and tourist visas, which would be the largest single revocation in American history.

What that already costs showed up in Bremen. Leiphone's report from IJCAI, the big AI conference that closed Saturday: mainland authors submitted over 4,000 papers, more than 60% of the total, and had 541 accepted — a 13.3% rate against Austria's 32.8%.

Chinese registrations came first at 669. The hall did not match. Program vice-chair Bo An, who chairs the programme next year, told Leiphone that strict refusals at the Guangzhou consulate kept many attendees out. Organisers are already discussing scheduling accommodations for Kyoto next August. Note also that CCF downgraded the conference from A-class to B and submissions rose anyway.

Go deeper on this section: ClaudeChatGPT

Thomson Reuters built its model on Qwen

Two Western legal-AI incumbents shipped on Chinese open weights this week. Leiphone leads with Alibaba: Thomson Reuters launched Thomson-1, built on Qwen3.5, to cut dependence on Anthropic's Claude, starting with table analysis and document review. Separately, Harvey — the American legal-AI company, valued around $11bn — post-trained Kimi K3, Moonshot's open-weight model, into a legal model called Tenet.

The Western write-ups bury the Chinese part. Thomson Reuters' own launch described only a strong open-source foundation and named its base model Snowdon, built with Imperial College London; the Qwen lineage came from the chief technology officer afterwards, with emphasis on months of realignment to be "ethically and politically de-biased". Cost: about $40m over two years.

Same fact, two stories. Chinese tech press treats it as a scoreboard entry — Alibaba claims 460-plus open models and three billion downloads, and cites Pinterest's chief executive putting Qwen's cost at under 8% of comparable closed models. American coverage treats the provenance as a risk to be managed, with Senator Tom Cotton warning about backdoors and Airbnb pushing back. Nobody disputes the mechanism: the cheapest good base weights are Chinese.

Go deeper on this section: ClaudeChatGPT

Memory is now bigger than semiconductors were

THE NUMBER: Gartner puts the 2026 memory market at $837.3bn — larger than the entire semiconductor industry's 2025 revenue of $809bn. DRAM up 246.6% in a year, NAND up 371.9%, memory's share of chip revenue doubling from 27% to 54%.

The capex consequence is already booked. TrendForce has DRAM and NAND going from 47% of major cloud providers' capital spending this year to 68% in 2027, with server memory contract prices up 64% in the second half of 2025 and 270% again this year. Nvidia has told server makers its flagship AI systems cost about 17% more.

Which is why Xiaomi's second chip matters more than its first. The Xring O3 scored 5.22 million on AnTuTu and beat Apple's cores on single-thread — 917 points on Hacker News, all about CPU. GeekPark's read is better: the score is fireworks, the bandwidth is the signal.

The O100 stacks memory wafers onto the compute wafer. Two AI-memory layers bonded to an NPU die at a 1.4-micron pitch, yielding 1.22TB/s — sixteen times the 76.8GB/s of LPDDR5X in today's flagships — and 330 tokens a second locally with a five-value quantised MiMo model that saves another 30% of bandwidth. Lei Jun has spent 21bn yuan (~$2.96bn) over five years to get here.

Go deeper on this section: ClaudeChatGPT

Everyone is a guest on somebody's show

China is arguing about interview surplus. Jiemian's essay opens on a video blogger posting on 21 August that he does not mind being interviewed, he minds living in an era of surplus meaning — everyone with a show, everyone a guest, everyone obliged to distil life into quotable lines.

The supply numbers explain the fatigue. Bilibili submissions of video podcasts over two hours are up nearly ninefold this year; single-day watch time passed 100 million minutes on 13 May; more than half run past 30 minutes. Veteran host Chen Luyu recorded six and a half hours with interviewer Jiang Sida and aired three hours fifty-five.

Meanwhile Douyin published its enforcement ledger. In one month it actioned 1.4534 million videos and 597,500 accounts for malicious commercial content — including 59 accounts compiling negative material on an unnamed famous technology brand, from accident footage to share-price moves, and 236 fabricating claims about a domestic beverage brand by wrapping them in nationalist sentiment.

The tell is the vocabulary. Rumour-mongering, commercial defamation, black-and-grey industry — a full taxonomy for protecting corporate reputation, deployed at a scale ByteDance has never announced for anything else.

Go deeper on this section: ClaudeChatGPT

Threads we are pulling

  • I gave the 1500m to the wrong robot, confirmed. IT Home now records the win for Beijing Humanoid's Tiangong Ultra at 2:21.63, not Honor's Lightning. Today the hundred metres fell twice: Lightning ran 8.94, then Tiangong ran 8.86. Scale of the event: 666 teams, 2,056 robots, 16 countries, 51 disciplines, 1,301 matches.
  • Doubao Work shipped on schedule. Yesterday I noted ByteDance folding Trae and Coze into Doubao with a work suite due within the week. It launched Tuesday — desktop app, cloud PC for long-running jobs, 30 days free, inheriting the user's Feishu permissions. GeekPark names the trigger: Tencent's WorkBuddy reached million-scale daily users and 20.97 million desktop visits in June.
  • Alibaba's management bought their own placement. After Sunday's HK$80bn raise, chairman Joe Tsai took 720,000 shares near HK$112 and chief executive Eddie Wu 350,000 near HK$111.6 — about HK$120m (~$15m) between them.
  • Xpeng set the embodied-AI record. Its robotics arm raised over $900m at a post-money valuation above $6.3bn, IDG leading, Tencent and Alibaba strategic — with a seven-year IPO deadline or investors can force a buyback at 8% compound. Unitree meanwhile fell 10.31% to 243.9bn yuan (~$34bn), passed by Figure AI.
  • The formaldehyde cabbage has a sequel. A pest-control firm in Xiamen allegedly sprayed dichlorvos, an organophosphate insecticide, in dozens of chain restaurants — on tableware and tables. A district joint investigation group seized its stock and closed its premises. Two food-poisoning scandals, two task forces, three days.

Go deeper on this section: ClaudeChatGPT

The river · 34

ResNet co-author Ren Shaoqing starts a humanoid-robotics company

# QbitAI

Ren Shaoqing, Nio’s autonomous-driving chief, has reportedly founded a physical-AI and humanoid-robotics company while Nio invests and collaborates with it.

More
  • The reported structure is deliberately ambiguous: Ren remains at Nio, the new company’s name is undisclosed and its unicorn-level valuation has not been independently detailed.
  • Nio’s strategic rationale is that autonomous driving and robotics share perception, prediction, planning, control, world models and reinforcement learning.
  • Ren’s prior work on ResNet and Faster R-CNN gives the venture unusual research credibility, while his Nio role supplies an industrial path into sensors, manufacturing and closed-loop deployment.
  • The move follows a broader Chinese pattern of automakers using robotics spinouts to attract capital and separate long-horizon experimentation from the core vehicle business.
  • The important test is whether Nio commits people, compute and manufacturing capacity at startup scale; a nominal strategic investment would not prove a serious robotics push.
Open at QbitAI Ask Claude ChatGPT

A Chinese legal drama revives the question of what lawyers are for

# Pekingnology

The state-produced drama The Pillar has prompted legal scholar Luo Xiang to argue that a justice system is tested by its treatment of lawyers defending unpopular defendants.

More
  • The series revisits real cases from China’s legal reconstruction between 1979 and 1997, including an episode where defense lawyers were arrested after challenging weak evidence.
  • Its production by the state broadcaster and top judicial and procuratorial institutions makes the subject notable: official media is presenting historical abuses as part of a lesson about legal development.
  • Luo’s intervention shifts the discussion from nostalgia to institutional principle, arguing that defense lawyers restrain state power rather than merely obstruct punishment.
  • The drama’s popularity shows how historical legal narratives can create a socially acceptable route for discussing due process and the dangers of public certainty.
  • The project does not demonstrate a present-day expansion of defense rights; its importance lies in the boundaries of what the current system is willing to remember and debate.
  • The strongest future signal would be whether the drama prompts sustained discussion of current defense practice rather than remaining a safely historical morality tale.
Open at Pekingnology Ask Claude ChatGPT

Chinese researchers dominated IJCAI submissions but struggled to attend

# Leiphone

IJCAI 2026 drew more than 4,000 Chinese submissions and 669 Chinese registrations, yet visa barriers left far fewer Chinese faces at the German venue.

More
  • China accounted for more than 60% of submissions and 541 accepted papers, while its reported acceptance rate was about 13.3%, below several smaller European delegations.
  • The gap between registrations and physical attendance exposes a practical geopolitical constraint on scientific exchange: institutional participation can look strong in data while researchers are absent from the room.
  • The report says China’s CCF downgrade of IJCAI from an A- to a B-level venue did not reduce submissions, suggesting that field relevance may matter more than domestic ranking labels.
  • A new primary-paper fee policy increased single-paper submissions from 25% to 40%, indicating that economic incentives can reshape publication behavior and reduce extreme multi-paper concentration.
  • The conference’s field mix still favored machine learning, vision, data mining and NLP, while symbolic reasoning, agents, planning and robotics remained smaller but identity-defining areas.
  • The organizers are considering scheduling changes and outreach for Chinese participants at IJCAI 2027 in Kyoto, a sign that access logistics are becoming part of conference governance.
  • The broader implication is that visa policy and domestic academic incentives can shape the geography of AI knowledge production as much as research quality or funding.
Open at Leiphone Ask Claude ChatGPT

Dobot reports embodied-AI revenue up more than 20 times

# Leiphone

Chinese robot maker Dobot reported first-half revenue of 320 million yuan, roughly $45M, up 106.6%, with embodied-AI revenue reaching 45 million yuan, roughly $6.3M.

More
  • Dobot’s core collaborative-robot business gives it an industrial distribution base: the company says it has shipped more than 100,000 units to over 100 countries and serves more than 80 Fortune 500 manufacturers.
  • Embodied-AI revenue grew more than 20 times and reached 14.3% of total revenue, with more than 60 million yuan (roughly $8.5M) in orders on hand.
  • The company says its Atom humanoid robot can operate for 14 hours a day in a cinema, selling more than 1,000 cups of popcorn daily without human assistance; this is a company-reported deployment claim.
  • Its DobotWAM model combines 3D geometric perception with motion constraints and is paired with arms, humanoids, wheeled and legged platforms in a one-brain, many-body architecture.
  • The strategic advantage is industrial feedback: real automotive, battery and logistics lines provide physical data that simulation and exhibition demos do not.
  • The result is one of the clearer tests of whether embodied AI can become an extension of an existing automation business rather than a capital-intensive standalone bet.
  • The main caveat is mix and margin: embodied AI remains a minority of revenue, so commercial leadership will depend on repeating deployments without eroding the profitable collaborative-robot base.
Open at Leiphone Ask Claude ChatGPT

Xi’s anti-corruption campaign is accelerating after 13 years

# SCMP

A review found 65 investigations of vice-ministerial officials or above in 2025, versus 19 in 2013, while the disciplinary campaign continues to expand under Xi Jinping.

More
  • The figures point to institutionalization rather than a short campaign: China’s top disciplinary body is still opening high-level cases more than a decade after Xi took office.
  • The supplied review says there were 65 probes of civilian officials at or above vice-ministerial rank last year, compared with 19 in 2013.
  • The CCDI had already opened 46 investigations this year by the report’s cutoff, versus 39 in the comparable prior period cited in the source.
  • At this scale, anti-corruption enforcement is both governance maintenance and a tool for political discipline, reshaping incentives across ministries, provinces and state firms.
  • The campaign can improve compliance and remove entrenched rent-seeking, but repeated purges also create risk aversion, policy paralysis and uncertainty for officials making discretionary decisions.
  • The key distinction to watch is whether enforcement is paired with durable procurement, asset-disclosure and oversight reforms or remains dependent on campaign-style discipline.
Open at SCMP Ask Claude ChatGPT

China seeks public reports of corruption in Rocket Force procurement

# SCMP

The PLA has opened a public tip channel for suspected procurement violations involving suppliers, bidders, evaluators and Rocket Force personnel as its corruption probe continues.

More
  • Inviting reports from vendors and tendering agencies widens the investigation beyond military officers to the commercial ecosystem that supplies the strategic missile branch.
  • The move implies that Beijing sees procurement integrity as a systemic control problem, not merely a few corrupt individuals.
  • The Rocket Force is especially sensitive because procurement failures can directly affect readiness, modernization and confidence in the nuclear deterrent.
  • A public reporting mechanism also gives investigators a way to surface local knowledge without announcing the full target list or investigative chain.
  • The notice does not establish that specific weapons programs failed or identify senior suspects, so the signal is about institutional scrutiny rather than proven operational damage.
  • What would make this materially more serious is evidence of procurement-driven readiness failures or a widening purge into other strategic commands.
Open at SCMP Ask Claude ChatGPT

China drafts law to extend its anti-corruption reach across borders

# SCMP

China's legislature began reviewing a draft Anti-Cross-Border Corruption Law, signaling that Beijing wants its anti-graft regime to cover overseas conduct as well as domestic officials.

More
  • The draft received its first reading at a session of the Standing Committee of the National People's Congress, but the text was not public in the supplied report.
  • The move would extend the legal architecture around corruption beyond China's borders, potentially covering cross-border transactions, foreign assets and conduct involving Chinese entities; the precise scope remains unclear.
  • It fits a broader pattern in which Beijing converts political campaigns into formal statutes, giving enforcement agencies a more durable basis and a wider jurisdictional narrative.
  • For multinational companies, the important issue is not the headline label but whether the law creates new disclosure, investigation or cooperation exposure for Chinese subsidiaries and counterparties.
  • The absence of the draft text means no responsible conclusion can yet be drawn about penalties, jurisdiction, evidence standards or interaction with foreign anti-bribery laws.
  • Chinese coverage presents the review as a major institutional initiative rather than a routine legislative item, consistent with the centrality of discipline enforcement in current governance.
  • Its significance will rise sharply if the final law reaches overseas personnel, intermediaries or assets directly; otherwise it may remain primarily a signaling instrument.
Open at SCMP Ask Claude ChatGPT

Chinese official calls for ‘normal stability’ with the United States

# SCMP

At a semi-official China-US dialogue, Party official Liu Haixing urged normal stability with controllable disagreements and a larger zone of shared interests.

More
  • Liu heads the Communist Party’s International Department, so the formulation is a useful signal of how Beijing wants the relationship described even if it is not a treaty commitment.
  • The phrase rejects an explicitly zero-sum framing while preserving room for disagreement, suggesting managed competition rather than a return to strategic trust.
  • Calling for a larger shared-interest pie is an invitation to compartmentalize economic and practical cooperation away from security disputes.
  • Track 1.5 talks are not the same as formal negotiations, but they can test language later used in official diplomacy.
  • The important uncertainty is whether this rhetoric is backed by concrete arrangements on trade, technology controls or military risk reduction.
Open at SCMP Ask Claude ChatGPT

China draft traffic law would make automakers answer for some automated-driving violations

# 36Kr

A Chinese roundup highlights a draft traffic-law chapter that assigns handling of violations during activated autonomous driving to vehicle makers or importers.

More
  • The draft distinguishes autonomous driving from assisted driving and says cars using only assistance remain governed by ordinary driver rules.
  • That split is strategically important: liability would follow the active control mode, creating a legal incentive for manufacturers to define, log and constrain operating domains precisely.
  • The proposal would move China beyond warning users that assistance is not autonomy, but it is still a draft and the supplied report gives no final text or implementation timeline.
  • The same roundup says SK Hynix closed its Chinese marketplace store and that Canada announced tariffs on US goods; those items are separate and do not change the significance of the traffic-law proposal.
  • For technology companies, the key unknown is how courts will handle mixed-control incidents where software is active but the human remains legally responsible.
Open at 36Kr Ask Claude ChatGPT

China’s muted response to North Korea’s missile launches is deliberate

# SCMP

After Wang Yi’s Seoul visit, North Korea launched more than 10 short-range missiles and Beijing responded cautiously, signaling that regime stability still outranks near-term denuclearization.

More
  • The launches came only hours after Wang concluded his first official South Korea trip in nearly five years, making the timing diplomatically awkward for Beijing.
  • Analysts cited in the report read China’s restraint as a choice to avoid destabilizing Pyongyang even when its behavior complicates China’s relations with Seoul and Tokyo.
  • This is consistent with a hierarchy in Chinese regional policy: preventing collapse, war or uncontrolled escalation can take precedence over enforcing maximal nonproliferation outcomes.
  • The muted response also preserves Beijing’s leverage with North Korea and avoids appearing to coordinate its Korea policy with Washington or Seoul.
  • For regional planners, the implication is that Chinese pressure on Pyongyang may be conditional and crisis-focused rather than a sustained campaign to eliminate its weapons program.
  • The signal would change if Beijing publicly condemned the launches, backed new sanctions or used senior-level diplomacy to impose costs.
Open at SCMP Ask Claude ChatGPT

Why China’s AI catch-up is moving so fast

# IT Home

Chinese coverage highlights a long-built talent network, open source and compute efficiency rather than a sudden breakthrough; the remaining bottleneck is advanced chips.

More
  • The report traces the current wave to Tsinghua’s long-running training network, especially the Yao Class and Tang Jie’s research group, with researchers moving between universities and startups.
  • Zhipu, Kimi and DeepSeek are presented as connected rather than isolated companies, with papers, open weights and engineering practices diffusing across teams.
  • The reported gap with America’s leading models has narrowed to months, although that estimate is attributed to industry participants rather than an independent benchmark.
  • DeepSeek’s multi-head latent attention and mixture-of-experts work illustrate the Chinese strategy of trading scarce compute for architectural efficiency.
  • The article says Chinese firms have historically invested less than one-fifth as much as US peers and often receive only about one-fifth as many high-end chips.
  • Talent returning from American labs and Chinese recruitment from overseas companies add another channel, but Liang Wenfeng’s reported view is that chips, not people, remain the decisive constraint.
Open at IT Home Ask Claude ChatGPT

China names 26 apps and SDKs accused of violating user rights

# IT Home

China’s industry ministry reported 26 apps and SDKs for illegal or improper collection and use of personal information, escalating a broader privacy-protection campaign.

More
  • The action is part of a joint 2026 personal-information protection program involving the cyberspace, industry and public-security authorities.
  • The ministry says the named software must rectify the violations and may face further action if remediation is inadequate.
  • The legal basis includes China’s Personal Information Protection Law, Cybersecurity Law and telecom regulations, showing how privacy enforcement is being operationalized through app inspections.
  • SDKs are included alongside consumer apps, extending scrutiny to the embedded components that can collect data across many downstream products.
  • The campaign signals that compliance is being treated as an ecosystem property rather than only an issue for the app visible to the user.
Open at IT Home Ask Claude ChatGPT

Unitree shares plunge 45% after Shanghai debut frenzy

# Techmeme

Chinese coverage is absent while Reuters, surfaced by Techmeme, focuses on Unitree's 45% post-debut plunge and the resulting warning about a Chinese robotics bubble.

More
  • Reuters reportedly puts Unitree's valuation near $36 billion after the fall, down from about $66 billion after its shares rose more than fivefold on debut.
  • The market move does not prove that Unitree's products or revenue prospects changed; it does show how Shanghai's appetite for humanoid-robot exposure can detach public pricing from operating evidence.
  • For Chinese robotics, a sharp correction can be strategically important because listed peers use market value to recruit engineers, finance factories and persuade industrial customers.
  • The key test is whether Unitree can convert valuation into repeatable shipments, service revenue and defensible hardware or model margins.
  • Until such data appears, the stock is functioning as a sentiment thermometer for the sector rather than a clean measure of robot adoption.
Open at Techmeme Ask Claude ChatGPT

Shengshu Technology sets out a five-level roadmap for general world models

# AI QbitAI

Chinese company Shengshu Technology proposes a five-level path from world generation to interactive environments, physical action, autonomous agents and multi-agent orchestration.

More
  • The company's Motubrain model reportedly unifies visual understanding, world-state prediction and action trajectories through a mixture-of-transformers design that gives modalities dedicated parameters and shared attention.
  • The source claims a tenfold inference-speed improvement over Motus, adaptation to a new robot body with 50 to 100 human demonstrations, and a 96.1 score on RoboTwin 2.0.
  • The roadmap's first three levels have concrete products: Vidu for coherent video worlds, Vidu S1 for real-time interaction, and Motubrain for robot action.
  • The higher levels require active exploration, long-term memory, goal formation and coordination among robots, people, tools and digital agents—capabilities not established by a video demo.
  • The useful strategic distinction is between generating a plausible scene and maintaining a world model that can survive feedback, physical action and multiple agents.
Open at QbitAI Ask Claude ChatGPT

Chinese AI-for-science platform moves from tasks to full projects

# AI QbitAI

Zidong Taichu upgraded ScienceClaw with AutoProject, an agent engine that plans, executes, verifies and records long-running research projects instead of isolated tasks.

More
  • Project2Task converts a broad research goal into a dependency graph with parallel and serial branches, while reusing data and model assets across experiments.
  • TaskExecutor monitors experiments, handles failures, revises hypotheses and reruns work, replacing one-shot calls with a feedback loop better suited to nonlinear scientific investigation.
  • EviGraph records evidence across subtasks and checks whether experiments support hypotheses and whether final claims exceed the available evidence.
  • The source reports an ARCBenchML score of 0.865 versus 0.596 for the best baseline, but also gives a claim-support rate of only 0.38, underscoring that traceability remains incomplete.
  • The architecture resembles a research CI system: planning, execution, validation and artifact retention are coupled instead of leaving scientists to stitch together separate tools.
  • The company explicitly says autonomous research does not mean researcher-free research; human judgment remains the likely bottleneck for question selection and interpretation.
Open at QbitAI Ask Claude ChatGPT

Open-source DM0.5 tops a difficult embodied-robot benchmark

# AI QbitAI

Yuanli Lingji says its open-source DM0.5 model leads RoboDojo, with especially strong memory and recovery on long-horizon manipulation tasks.

More
  • The benchmark is described as unusually difficult: as of July, leading average success rates were only 8.8% in simulation and 12.8% on real robots.
  • DM0.5 reportedly achieved 100% on three randomized trials of a task requiring the robot to remember hidden block colors and reveal them in order.
  • The model combines a deliberative high-level system with a faster action-specialist system, allowing planning and fine-grained control to operate at different timescales.
  • Demonstrations include sub-centimeter block assembly, flexible-object cutting and conveyor-belt parcel separation at an average of three seconds per item with more than 99% accuracy.
  • The broader lesson is that memory, human demonstration and closed-loop correction may matter more than simply adding training hours when tasks require persistent state.
  • The claims come from promotional coverage and benchmark demonstrations; independent replication and real deployment data remain the missing evidence.
Open at QbitAI Ask Claude ChatGPT

China’s official textbook publisher partners with Xiaoyuan on an AI tutor

# AI QbitAI

The People’s Education audio-visual publisher and Xiaoyuan launched a textbook-based learning agent that combines approved curriculum content with a proprietary education model.

More
  • The product separates AI companionship from AI instruction: one mode encourages reading, speaking and engagement, while the other follows curriculum goals and builds a step-by-step learning path.
  • The system can turn textbook pages into interactive lessons, answer questions, assess recitation and adapt exercises using structured educational resources rather than generic web content.
  • This partnership shows how China’s education-AI market is likely to be shaped by authorized publishers, device makers and model providers working together under tighter content-safety expectations.
  • The companies say a pilot attracted more than 100,000 students and reported 33.6% higher Chinese-language knowledge mastery and 24% higher English mastery, but the figures are internal and lack a control group in the supplied text.
  • The policy context is important: a 2026 national AI-plus-education plan calls for intelligent learning companions and themed learning scenarios, giving the product an institutional route to scale.
Open at QbitAI Ask Claude ChatGPT

A Chinese startup tries to make forecasting a reusable system skill

# Signal QbitAI

DigClaw’s Rhizome framework placed three different base models in the FutureX top seven, arguing that search, causal reasoning and probability calibration can live outside model weights.

More
  • FutureX uses 59 unresolved real-world questions and settles them later, giving the system feedback that ordinary language-model benchmarks cannot provide.
  • Rhizome separates retrieval from reasoning, preserving the full evidence and tool-use trajectory before the outcome is known so errors can be attributed to search, timing, inference or calibration.
  • It aggregates probabilities in logit space, adjusts overconfidence using Brier scores and applies Platt scaling as results accumulate, treating calibration as a learned operational asset.
  • The cross-model result suggests that orchestration, evidence management and temporal discipline can transfer across base models, allowing institutions to trade off capability, cost and latency.
  • The evidence is still early: three leaderboard placements and a company technical report do not prove reliable forecasting in high-stakes markets or policy settings.
  • The most valuable asset may be the time-indexed record of predictions and outcomes, which cannot be recreated after an event is known.
Open at QbitAI Ask Claude ChatGPT

ByteDance consolidates its AI products around Doubao Work

# AI GeekPark

ByteDance folded TRAE and Coze teams into Doubao while keeping the coding IDE and CLI distinct, positioning Doubao Work as its unified office-agent front end.

Also: Leiphone, 36Kr

More
  • The organizational move resolves overlapping product directions: TRAE Work, Coze and Doubao are converging on task execution, while TRAE IDE and CLI remain a programming line under the Doubao brand.
  • Recent features include remote computer control, virtual desktops, a skill store, connectors and more than 200 tools, creating a full agent harness rather than a chatbot.
  • Feishu adds enterprise identity, permissions, messages, documents and meetings, giving the agent context that ordinary prompt-based products lack.
  • The move appears responsive to competition from Tencent’s WorkBuddy and Alibaba’s office-agent efforts, suggesting that Chinese platforms are now racing to package agents for ordinary workers rather than only developers.
  • The strategic question is whether consolidation improves reliability and distribution or merely hides duplicated capabilities behind one brand.
Open at GeekPark Ask Claude ChatGPT

HarmonyOS 7 turns app icons into callable services

# AI GeekPark

Huawei’s HarmonyOS 7 lets system AI discover app capabilities, call them through skills or agents and keep sensitive workflows on-device.

More
  • The architectural shift is from app-launch metrics to service invocation: a user can ask the system to perform an action without finding the app icon.
  • Huawei distinguishes simple parameter-in/result-out skills from multi-turn agents, while an on-device agent-to-agent interface lets banks keep conversation data on the handset.
  • The system exposes capabilities such as image super-resolution and semantic image search through platform APIs, potentially reducing each developer’s need to ship and optimize separate models.
  • The article says the system can preserve enterprise permissions and return generated work to the organization rather than trapping it in a personal chat, a design that resembles an operating system for agents.
  • This is a platform-control play: Huawei is trying to own distribution, context and privacy boundaries while developers retain domain logic and user-facing services.
  • The limitation is reach. These interfaces exist only on HarmonyOS devices, and the supplied figures show agent usage still far below total app usage.
Open at GeekPark Ask Claude ChatGPT

Beijing combines economic retaliation with a continuing Rocket Force purge

# Policy Sinocism

Sinocism highlights Beijing’s defense of cooperation with Iran, cadre accountability for environmental harm and ongoing procurement scrutiny inside the Rocket Force.

More
  • The foreign ministry statement says China’s cooperation with Iran is lawful and that Beijing will protect its rights and interests, a deliberately broad response that leaves room for escalation.
  • The issue also points to articles about the Zhong Caiwen economic narrative and official accountability for environmental damage, suggesting that economic resilience and cadre discipline are being framed together.
  • The Rocket Force thread is the most consequential: procurement corruption is being treated as a continuing institutional problem rather than a closed scandal.
  • Because this is a newsletter digest, the supplied text is selective and does not establish the full scope or seniority of the officials involved.
Open at Sinocism Ask Claude ChatGPT

Anonymous Chinese model launches into overseas developer workflows

# AI GeekPark

Chinese coverage frames the anonymous Ox Alpha model as a deliberate overseas testing strategy: free access, long context and agent integration let developers test it before its maker claims it.

More
  • Ox Alpha reportedly offers a one-million-token context window, image and video input and tool use through OpenRouter, OpenCode and related coding agents.
  • The community suspects a relationship with Zhipu’s GLM line based on tokenizer behavior, error messages and output habits, but the company has not confirmed the identity.
  • Early DeepSWE tests looked strong on only ten tasks, while a larger community retest fell to about 63%; those results are not audited benchmarks and may reflect gateway failures as well as model quality.
  • The anonymous-release pattern follows earlier cases in which Alibaba, Xiaomi and Ant teams later acknowledged anonymous model aliases, suggesting that overseas developer platforms are becoming informal launch channels for Chinese models.
  • This is a different go-to-market model from a domestic press conference: developers encounter the model inside an existing toolchain, generate the first wave of evaluation and sometimes identify the supplier themselves.
  • The unresolved issue is trust. A model can win attention with a small free test while leaving developers uncertain about provenance, data handling, uptime and future pricing.
Open at GeekPark Ask Claude ChatGPT

Spatial intelligence needs a state ledger, not just a rendered world

# Analysis Leiphone

A Chinese analysis of spatial intelligence argues that generated 3D worlds need persistent, verifiable state so objects, rules and consequences survive continued interaction.

More
  • The proposed system separates a renderer that shows the world, a simulator that updates it and a planner that chooses actions; the shared state layer is the architectural hinge.
  • A usable world must track semantic identity, geometry and runtime state together, so a moved table remains the same table and a changed door updates navigation constraints.
  • This makes world generation resemble compilation and continuous integration: natural language becomes constraints, assets fill the scene, solvers enforce geometry and validators decide whether a new state can be committed.
  • The article distinguishes editing the world from acting inside it, implying different permissions, dependencies and rollback semantics for each operation.
  • Its recursive self-improvement loop stores state, goals, actions, verification and feedback, then generates harder cases and gates new versions against regressions.
  • Examples involving Kimi and other model systems show why Chinese developers are using generative models as a route into programmable 3D environments, but the article’s broader technical thesis extends beyond China.
  • The key product metrics become time to first valid scene, cost per valid result and the ability to make local changes without recomputing or corrupting the whole world.
Open at Leiphone Ask Claude ChatGPT

Xi’s anti-corruption campaign extends beyond serving officials

# Policy SCMP

An SCMP data project says China’s anti-graft campaign is targeting officials outside Beijing and long after retirement, broadening the perceived exposure window.

More
  • The report’s central finding is institutional rather than case-specific: senior officials cannot assume that leaving office or working outside the capital ends their corruption risk.
  • That changes the incentive structure for cadres, because local postings and post-retirement networks remain within the reach of investigators.
  • The story places the campaign within Xi’s signature anti-corruption drive, where discipline enforcement also functions as a mechanism of political control.
  • The supplied excerpt does not identify new arrests or a quantified target list, so it is better read as analysis of enforcement direction than as a fresh purge announcement.
Open at SCMP Ask Claude ChatGPT

PhanthyMotus seeks a common software layer for Chinese robots

# AI QbitAI

Phanthy says more than ten robot makers are co-building its open embodied-agent platform, which claims 15 hardware adapters and 200 standardized driver modules.

More
  • The proposed middleware separates hardware, algorithms and applications so developers do not rewrite control and skill code for every robot body.
  • The project claims more than 600,000 lines of contributed code across humanoids, quadrupeds, wheeled robots, arms and drones, but the source is company-sponsored and offers no independent code-quality review.
  • Its developer incentive program includes remote access to hardware and free compute tokens, an attempt to solve the physical-access problem that limits embodied-AI experimentation.
  • The real test is governance: a hardware-neutral layer must maintain stable interfaces while vendors still differentiate their bodies, safety systems and commercial data.
  • The target of 50 adapters and 10,000 shared skills by the end of 2026 is an ecosystem ambition, not evidence that cross-body zero-cost reuse already works in production.
Open at QbitAI Ask Claude ChatGPT

China’s tech roundup pairs chip bets, AI consolidation and robot valuations

# Signal Leiphone

A Chinese roundup says Xiaomi unveiled a three-chip domestic portfolio, ByteDance consolidated office agents and Unitree lost the top humanoid valuation spot to Figure AI.

More
  • Xiaomi says it has invested more than 21 billion yuan, roughly $3.0B, and built a nearly 3,000-person chip team; its O3, O100 and D100 processors span phones, computing and driving.
  • The D100 is described as a domestic 3nm driving-AI chip, while the O3 is already in volume production and the other chips are planned for commercial use in 2027.
  • ByteDance’s TRAE and Coze integration into Doubao and the launch of Doubao Work show a major Chinese platform narrowing its AI-product architecture around one work-agent brand.
  • The roundup says Unitree’s market value fell to about 243.9 billion yuan, roughly $34B, below Figure AI’s reported $39B valuation; both numbers are market or reported valuation figures, not equivalent operating metrics.
  • It also reports Nio backing a new embodied-AI company founded by its autonomous-driving chief, illustrating how Chinese automakers are exporting talent and data into a broader physical-AI stack.
  • Because this is a roundup, each item needs separate verification; its value is the pattern of capital, talent and product consolidation across China’s technology sector.
Open at Leiphone Ask Claude ChatGPT

China’s 2026 Science Exploration Award names 50 young researchers

# Culture QbitAI

The eighth Science Exploration Award selected 50 researchers across ten fields and 13 cities, with each recipient receiving 3 million yuan, roughly $423,000, over five years.

More
  • The award’s geographic spread includes institutions and cities appearing for the first time, suggesting that China’s research capacity is broadening beyond Beijing and a small group of traditional centers.
  • Its scientist-led selection model funds future exploration rather than simply rewarding completed work, offering a private-philanthropy complement to state research programs.
  • Recipients work across mathematics, materials, medicine, information technology, energy, manufacturing and other fields, showing an effort to treat basic science and frontier engineering as one talent pipeline.
  • The award has supported 347 scientists since 2019, according to the source, but the article does not track later research outcomes.
Open at QbitAI Ask Claude ChatGPT

Reuters parent builds a model on Alibaba’s Qwen

# AI Leiphone

Thomson Reuters reportedly built Thomson-1 on Alibaba’s open-weight Qwen3.5 to reduce dependence on Claude and target document review and spreadsheet analysis.

More
  • The move is strategically notable because a major Western information company is using a Chinese open-weight model as a base for a specialized internal system.
  • Open weights let an enterprise control deployment, adaptation and cost, but the source does not disclose Thomson-1’s evaluation results, data governance or production scale.
  • The model is intended to take over some tasks previously handled by Anthropic’s Claude, suggesting that open models are competing on total operating economics rather than only benchmark scores.
  • Chinese coverage emphasizes Qwen’s global downloads and derivative ecosystem, but those figures are company or community claims and do not establish enterprise reliability.
Open at Leiphone Ask Claude ChatGPT

Chinese embodied-AI research targets the gap between demos and physical generalization

# AI Leiphone

A Leiphone review of IJCAI 2026 work says leading VLA systems can exceed 90% on standard tests yet fall below 21% on shifted tasks without stronger control and representation methods.

More
  • One cited study found models reaching below 21% success on novel task arrangements, then raised one system from 9% to 83% by interpolating internal task-semantic representations without retraining the full model.
  • Another method adapted a visual foundation model by tuning about 4 million parameters instead of 632 million, cutting training time while preserving most reported segmentation performance.
  • The common theme is that a capable policy can still memorize coordinates, appearances or benchmark structure instead of learning transferable physical concepts.
  • These techniques complement rather than replace large models: the high-level model supplies semantics, while specialized perception, control and adaptation layers handle embodiment and latency.
  • The figures come through a conference review and paper summaries, not a common independent evaluation, so comparisons across tasks should be treated cautiously.
Open at Leiphone Ask Claude ChatGPT

SenseNova U1.5 Lite open-sources an 8B image model

# AI QbitAI

SenseTime released the formal version of its 8B image model, claiming stronger layout control and local editing while approaching closed systems on selected tasks.

More
  • The demonstrations emphasize multi-image composition, structured infographics and localized edits that preserve scene geometry rather than regenerating the whole image.
  • An 8B model is small enough to make local deployment plausible, but the report does not specify hardware requirements, license terms or independent comparisons.
  • The source claims parity with GPT-Image-2 on complex layout and editing, yet all examples and evaluations are presented by the model maker’s promotional coverage.
  • The technical direction is useful: controllable editing and preservation of spatial relationships matter more to production workflows than attractive first-pass images.
Open at QbitAI Ask Claude ChatGPT

Chinese J-16s train with Egyptian Rafales amid regional war

# Policy SCMP

China’s air force deployed J-16 fighters for combat training with Egypt’s Rafales in a joint exercise as the United States-Iran conflict reshapes Middle East security calculations.

More
  • The exercise is the second China-Egypt Eagles of Civilization drill and is expected to run into early next month.
  • The choice to train J-16s alongside Rafales gives both sides experience operating across different aircraft, doctrines and logistics ecosystems.
  • The timing matters: China is signaling that it can maintain an operational presence and partnerships in a region where US military action is dominating the security agenda.
  • This is not evidence of a formal Chinese alliance with Egypt, but it expands practical military familiarity and diplomatic visibility.
  • The supplied text does not state the exercise’s scenario, weapons employment or whether the deployment has a deterrent purpose beyond training.
Open at SCMP Ask Claude ChatGPT

Chinese researcher argues AI’s next scaling edge is data efficiency

# AI Leiphone

At IJCAI 2026, BIMSA researcher Wang Yaqing argued that structured priors and data-efficient agent learning are necessary when more data and compute cannot solve the problem.

More
  • The research agenda targets domains such as drug discovery and cold-start recommendation, where useful labeled examples are intrinsically scarce rather than merely expensive to collect.
  • Wang’s framework links few-shot learning, meta-learning, in-context learning and agent systems, treating in-context adaptation as a form of data-dependent meta-learning under the studied assumptions.
  • The proposed Data-Efficient Agentic Learning approach combines experience enhancement, architecture design and learning strategies to address long-horizon interaction with limited feedback.
  • The technical takeaway is that agent scaling may depend on better representations of prior structure and experience, not only larger models or longer trajectories.
  • The supplied article is a conference interpretation rather than a peer-reviewed result report, so its broad claims should be separated from the formal results of the underlying work.
Open at Leiphone Ask Claude ChatGPT

Washington may force countries to choose between US and Chinese AI blocs

# Policy SCMP

A reported US draft letter would exclude countries joining Beijing’s AI cooperation framework from a US-led coalition, turning model access into alliance politics.

More
  • The draft language frames participation in both systems as incompatible, making standards, infrastructure and research partnerships instruments of geopolitical alignment.
  • The story is not a confirmed policy: the supplied report describes a draft and an official summary rather than a published decision.
  • For countries in the middle, the practical issue is whether they can use Chinese models, data centers or standards while retaining access to US chips, cloud services and capital.
  • The Chinese-facing significance is that Beijing’s proposed AI cooperation framework is being treated by Washington as a competing bloc, not merely as an open technical forum.
  • This could accelerate technological bifurcation even where governments would prefer vendor diversity and nonalignment.
  • The policy matters most if the final rules attach concrete export, investment or cloud-access penalties to participation rather than serving as diplomatic pressure.
Open at SCMP Ask Claude ChatGPT

Brazil weighs processing capacity after Washington admits China’s minerals lead

# Policy SCMP

Brazil’s mining minister says US officials conceded China had won the critical-minerals race, sharpening the question of whether Brazil can build processing power rather than just export ore.

More
  • The report makes a distinction that matters for industrial strategy: mineral reserves are not the same as refining, chemical processing, equipment and supply-chain control.
  • Brazil is presented as a potential swing supplier, but its ability to challenge China depends on infrastructure, financing, technology and environmental permitting.
  • The claim about the White House conversation is attributed to Brazil’s minister and was made in an unscripted speech, so it should not be treated as a formal US policy admission.
  • For China, the story reinforces the value of downstream processing and industrial ecosystems, not merely access to mines.
Open at SCMP Ask Claude ChatGPT