Feng Rui was detained after helping publish a wide-ranging review of Chinese labor issues; reports describe sustained questioning and censorship of coverage about his case.
More
- China Digital Times reports that Feng was detained in Guangzhou on September 14 on a charge commonly used against activists, following years of documenting wage arrears, workplace accidents, strikes and other labor disputes.
- The outlet says his team’s 2025 review covered 15 subjects, including informal work, overwork, social security, AI and automation, and women’s labor.
- The report says the review was censored online and the team’s social-media account suspended before Feng and his family faced pressure from local authorities.
- China Digital Times has archived reports about the detention, several of which it says were deleted from their original platforms; this makes the information environment part of the story, not just the arrest.
- The case shows how labor reporting can be treated as a political challenge when it documents patterns of workplace harm rather than isolated disputes.
- The account is based on reporting and archived online material; the available text does not include a response from authorities or a full account of the legal case.
A draft Chinese export-control catalogue covers 23 AI technologies alongside biomedical and photovoltaic areas, signaling a wider effort to manage outbound know-how.
More
- The newsletter reports that the commerce and science ministries proposed adding technologies in AI, biomedicine and photovoltaics to China’s restricted or prohibited export catalogue.
- The proposed AI coverage includes large models, agents, chip design, computing clusters, and selected cybersecurity and information-processing capabilities.
- The draft reportedly excludes technology already public and says some new open releases could receive automatic licensing; API and SaaS delivery generally fall outside large-model export management.
- The registration requirements for covered exports, including transfers already under way, took effect immediately, while comments on the draft are due November 15.
- The proposal also offers additional facilitation for technology exports that support cooperation among members of the planned World Artificial Intelligence Cooperation Organization.
- The practical boundary remains uncertain: the catalogue is still a draft, and its treatment of future model releases could change before finalization.
Taiwan’s coastguard says 16 mainland vessels operated across several areas in a coordinated one-day push, a pattern experts view as a possible escalation in Beijing’s maritime pressure.
More
- The vessels reportedly included 12 coastguard and research ships and four fishing boats, operating simultaneously near Hualien, Pingtung, Kaohsiung and the Dongsha Islands.
- Taiwan’s coastguard said it warded them off after they entered waters Taiwan designates as restricted; the excerpt does not give Beijing’s account of the operation.
- The spread across multiple maritime zones matters more than any one encounter: it can force Taiwan to divide surveillance and response resources across a wider area.
- The report describes the operation as a possible shift in pressure tactics, but the supplied excerpt does not establish the vessels’ precise missions or command arrangements.
- Beijing’s maritime pressure on Taiwan often involves state and civilian vessels, making attribution and escalation management difficult; that background should not be mistaken for proof of a specific intent here.
OpenCode says an obsolete system was accessed without authorization and API keys may have leaked; Chinese coverage focuses on exposed credentials and questioned the company’s link to model distillation.
More
- The report says the affected credentials were third-party model API keys stored in OpenCode Console, a cloud feature for syncing or sharing provider access.
- OpenCode reportedly said 1.4% of Console users were affected and connected the incident to a large model-distillation attack; developers questioned why that explanation fit a breach involving a legacy system and exposed keys.
- The article describes users reporting unauthorized bills from Anthropic, OpenAI and Google, including one account with more than $1,700 in charges in a day; those figures are reported user claims, not independently verified totals.
- Centralizing provider credentials can simplify team access, but it also creates a high-value secret store: a compromise may let an attacker call providers directly rather than operate through the application’s own quota controls.
- The report says some providers revoked keys and that attacks continued for 48 hours; it does not include a complete forensic account or explain the exact access path.
- The practical signal is about cloud control-plane design: minimize stored privileges, isolate secrets from retired systems, and make usage anomalies visible at the provider boundary.
- Whether the incident reflects a broader compromise or a bounded legacy-system failure depends on the investigation’s eventual findings, which the supplied text does not provide.
China's draft rules would leave most drug-licensing deals and contract research outside controls while restricting transfers of certain biotech platforms; the distinction is between exporting products and giving away reusable capabilities.
More
- A Commerce Ministry and science ministry draft adds technologies in AI, biomedicine and photovoltaics to the restricted-export catalogue, with prior registration applying during the consultation period.
- For biomedicine, the draft targets ownership transfers of certain cell and gene therapy, small-nucleic-acid drug and pathogen-related platforms; it also restricts some new-target discovery and validation technologies.
- The draft expressly excludes licensing and contract research or manufacturing services, and says processes needed to develop or manufacture a specific drug are outside the platform controls.
- That boundary preserves the common license-out model, in which a company licenses overseas rights to a drug candidate while retaining ownership of the underlying platform; it is not a blanket block on biotech exports.
- Compared with a version circulated earlier in the year, the draft drops antibody and small-molecule drug categories, a narrowing that signals an attempt to balance technology security with a still-important source of overseas deal revenue.
- The strategic target is reusable know-how rather than each individual medicine: keeping platform ownership in China could strengthen domestic firms' leverage in future partnerships, but the rules' final scope and enforcement remain unsettled.
- The supplied report says 121 China-to-overseas biotech transactions occurred in the first half of 2026 and estimates only a small fraction could be affected; that estimate is the article's analysis, not a regulator's impact assessment.
Ulanqab is using cool weather, low-cost power and fast links to Beijing to attract AI data centers, bringing new technical jobs to a region better known for potatoes and wind power.
More
- The article describes Ulanqab as having signed 109 computing projects and put 172,000 units of computing capacity into operation by August 2026; it says more than 95% is AI-oriented capacity.
- Local advantages are unusually concrete: a reported average temperature of 4.3 C, natural cooling for much of the year, power prices around 0.32 yuan per kWh (about 4.5 cents), and network latency to Beijing below 5 ms.
- The buildout links renewable generation to data-center demand, aiming to keep more value in Inner Mongolia instead of exporting electricity to other provinces.
- The shift also creates demand for electricians, HVAC technicians and other facilities specialists, not just software engineers; the article profiles a young returnee training for data-center operations.
- The same report cites a planned DeepSeek expansion of 1 GW, with some capacity expected in late 2027 or early 2028, and a local target of more than 8 billion yuan in computing-industry revenue in 2026.
- The investment claims are large: Ulanqab projects reportedly represent more than 500 billion yuan (about $70B) in total investment, against a city GDP of roughly 120 billion yuan; actual buildout and utilization will determine whether that pipeline becomes productive capacity.
- The story matters beyond AI infrastructure because it shows how China is trying to convert renewable-energy regions into service economies, though the article's headline figures are not independently substantiated in the supplied text.
Beijing is extending dual-use export controls to tunnel-boring machines and specialized equipment, presenting the move as nonproliferation and national-security policy while retaining a licensing channel for compliant trade.
Also:
Jiemian
More
- The controls cover complete tunnel-boring machines, key components and rebuilding technology, as well as certain walking excavators and high-altitude concrete sprayers.
- The scope reaches beyond finished machines to components and specialized production or hydraulic technologies, giving the policy leverage across parts of the supply chain.
- China says exports to sensitive destinations require permits; other destinations face registration and end-user commitments, with transfers onward to sensitive destinations also restricted.
- The Commerce Ministry describes the measures as consistent with international practice and says compliant applications can be approved. That framing emphasizes managed trade rather than a blanket embargo.
- The move shows how Beijing is applying dual-use controls to industrial equipment, not only to chips and other frequently discussed strategic technologies; the text does not identify specific countries or explain the security rationale for each category.
Investors reportedly cannot sell robot-company shares even at steep discounts, as IPO prospects and revenue quality come under scrutiny; the gap between headline valuations and exit prices is becoming harder to ignore.
More
- One investor reportedly offered shares in a humanoid-robot company at a valuation roughly half the company’s stated figure and still found no buyer.
- The article describes a reversal from the previous year, when investors competed for private stakes in leading robotics companies and secondary-market buyers were active.
- Market participants cited in the story say prospective listings face greater scrutiny of recurring revenue, real customer demand, narrowing losses and scalable shipments; whether this reflects formal regulatory guidance is disputed.
- Secondary-share prices can reveal liquidity and investor appetite more directly than new-round valuations, which may be set by fresh financing terms rather than cash exit opportunities.
- The reported market is bifurcating: stronger companies continue to attract capital while weaker firms face stalled sales, unpaid obligations or insolvency proceedings.
- The article attributes some inflated revenue to data-collection centers and subsidies, but the extent of such accounting and the eventual valuation impact remain uncertain.
A Tsinghua-led team reports a wearable microscope that records thousands of neurons in freely moving mice, removing a major constraint on natural-behavior neuroscience.
More
- The Nature paper describes Airscope, a wireless microscope weighing one gram that records neural activity at single-cell resolution across a six-millimeter field of view.
- The reported system uses four aspheric lenses and a two-megapixel sensor with on-device image compression, enabling wireless transmission without tethering the animal.
- The team says it recorded more than 8,500 neurons in one mouse at once, roughly 25 times the count of comparable devices.
- The technical tradeoff is unusually aggressive integration: the design aims to retain wide-field resolution while reducing the head-mounted mass that can alter animal behavior.
- Experiments included tracking multiple mice exploring a large habitat and following an animal through a long maze; the paper reports that cortical activity varied systematically with terrain.
- The result is a research instrument, not a demonstrated clinical device, but it could help researchers study neural activity during social and exploratory behavior under less artificial conditions.
The DHS has ordered civilian agencies to remove DeepSeek products and bar future purchases, turning US security concerns about a Chinese AI provider into a procurement restriction.
More
- The item reports an order covering federal civilian agencies, not a general ban on DeepSeek for US businesses or consumers.
- For Chinese AI providers, federal procurement is an important signal: security objections can constrain access to government customers even when the product remains commercially available elsewhere.
- The aggregator cites the Wall Street Journal; the supplied text gives no detail on the order's evidence, implementation scope, or any response from DeepSeek.
- The order's broader significance depends on whether other agencies or allied governments adopt similar restrictions.
A Tencent-led research team alternated verified experience extraction with policy refinement and raised a Qwen-based agent’s average score across terminal benchmarks.
Also:
TMTPost
More
- The report identifies Alibaba’s Qwen3.8-27B as the base model and describes a five-stage training sequence alternating recursive self-rewrite and recursive policy refinement.
- In the self-rewrite stage, the system selects successful task trajectories, turns them into reusable instructions and tests them in a fresh environment before retaining them.
- The policy-refinement stage uses reinforcement learning and recalibrates the value network to account for the agent’s changed behavior.
- The reported average across four Terminal-Bench suites increased from 20.68% to 50.58%; the study also evaluates long-horizon tasks and workflow reconstruction.
- The article cautions that a comparison favoring the experience-first route does not fully control for the extra supervision and compute used to discover and rewrite successful traces.
- This is a research result on a single model scale, not evidence that agents can improve indefinitely or without human-designed validation.
DeepSeek’s V4.1 Flash update leads an open-model web-development ranking and emphasizes low API costs, though the reported results are largely benchmark and product claims.
More
- The report says Arena ranked DeepSeek V4.1 Flash-1017 first among open models on its Code Arena web-development leaderboard, narrowly behind the overall leader.
- The article also cites DeepSWE results and a roughly 60% reduction in average task cost versus the prior version; those figures are benchmark claims rather than independent evidence of general coding ability.
- The model is presented as a sparse mixture-of-experts system with 552 billion total parameters and 8 billion activated for input and 16 billion for output, according to the report.
- Its price varies by cache status and demand period, with cached input particularly cheap; that makes workload mix and cache hit rate central to the real cost comparison.
- The outlet’s demos show the model generating a 3D scene and a browser game, but a handful of demonstrations cannot establish reliability on production software tasks.
MiniMax says its new coding and agent model reaches 200-plus tokens per second and cuts agent-task cost, positioning it as a direct low-cost rival to DeepSeek’s Flash model.
Also:
Hacker News
More
- The report describes M3.1 Flash as a lightweight coding and agent model with a one-million-token context window and prices of 1 yuan per million input tokens and 4 yuan per million output tokens.
- MiniMax says an aggressive sparse-attention design supports output above 200 tokens per second, with a reported median above 170 tokens per second.
- The outlet cites company benchmarks showing strengths in long-horizon agents and cross-application automation, while DeepSeek scores higher on one software-engineering benchmark.
- The comparison is useful for system design because it separates task success, decode speed, context length and cache pricing; benchmark leadership alone does not establish lower end-to-end cost for every workload.
- Both models compete in a segment where low-priced inference can expand usage, but the report’s claims are not a controlled independent comparison across identical serving stacks.
#
Culture China Digital Times
A translated Chinese social-media essay argues that foreign criticism of China is often treated as interference while flattering influencer content is amplified as positive publicity.
More
- The essay contrasts suspicion toward foreign coverage of domestic problems with enthusiasm for foreign visitors praising Chinese cities, food, shopping and technology.
- It points to tourism promotion by local authorities and argues that paid or invited influencer videos can be mistaken for spontaneous testimony.
- The author’s central complaint is that a foreign visitor’s praise can be used to override residents’ own accounts of everyday problems.
- The piece reflects an internal debate about propaganda and credibility: some commenters celebrate favorable foreign attention, while others ask who arranged and funded it.
China Telecom’s Xing4.0-29B-Intent combines intent classification with entity and parameter extraction, aiming to turn natural-language instructions into usable JSON for agents.
More
- The model is designed to return both an intent and structured fields, such as which device to change and which values to apply, rather than only selecting from fixed choices.
- The report says the model, training data and inference SDK were released, with a schema-constrained JSON output intended to reduce downstream parsing work.
- The outlet cites a zero-shot Micro-F1 score of 70.27 and latency below 150 milliseconds; those figures come from reported evaluations and the article’s own testing.
- The architectural distinction is between deciding what action is intended and extracting the arguments needed to execute it, a useful separation for tool-using systems.
- Whether it transfers to production depends on coverage of real user phrasing, schema changes and error handling when a model produces a plausible but unsafe parameter.
Recruiters describe selective hiring and steep pay premiums for foundation-model specialists, creating tension with incumbent employees as broader hiring contracts.
More
- The report says major internet companies are restricting headcount overall while still making exceptions for highly sought-after foundation-model researchers.
- Recruiters estimate that annual packages for new foundation-model graduates can start around 2 million yuan (about $280,000), with reported top offers of 4 million to 5 million yuan (about $560,000 to $700,000).
- For comparison, the article recalls historical top graduate offers of roughly 200,000 to 250,000 yuan (about $28,000 to $35,000) and a 600,000-yuan (about $85,000) ceiling in an earlier elite program.
- The report describes employee resentment over new hires earning more than higher-ranked staff, while some experienced engineers now treat the disparity as ordinary market pricing.
- Recruiters say the squeeze falls most heavily on application-side roles and mid-tier candidates; proven experience training very large models remains the scarce credential.
- These figures come from recruiter and employee accounts, not disclosed company compensation data, so they are a signal of hiring-market sentiment rather than a sector-wide pay census.
Both governments are trying to secure investment agreements before next month’s summit, while bargaining over leverage in extending a broader trade arrangement.
More
- The South China Morning Post reports that negotiators are working toward concrete investment deals ahead of a leaders’ meeting in Shenzhen next month.
- Sources describe intensive talks but say the sides still face hurdles in finalizing a package spanning institutional frameworks and economic agreements.
- The parallel contest over leverage in extending the broader trade arrangement suggests the investment package is part of a wider negotiation, not a standalone commercial push.
- The available report does not specify the deals under discussion or indicate that either side has reached agreement.
Conway Research’s Underdog runs on a user’s device and uses models based on Qwen, making privacy and local inference its answer to cloud-based personal agents.
More
- The product is designed to run locally and offline, with access to email, calendars, meeting notes and image generation while keeping personal context on the device.
- The report says its Saluki 27B model is based on Alibaba’s Qwen3.8-27B and that the company also offers a smaller model for faster local tasks.
- The startup claims its Husky inference engine substantially outperforms MLX on selected Mac tests, but those are company-reported results rather than independent comparisons.
- The product is still in a limited test phase; local execution may improve privacy, but it shifts compute, compatibility and update burdens to user hardware.
Trump says he is open to a three-way meeting in China about nuclear weapons, presenting a possible summit as his proposal rather than an agreed agenda.
More
- The South China Morning Post reports that Trump said he would accept a three-way meeting with Xi Jinping and Vladimir Putin during his planned China visit next month.
- Trump framed nuclear weapons and denuclearization as the agenda and said the proposal was his own; the report does not indicate that Xi or Putin accepted it.
- For Beijing, the proposal would put China at the center of a high-profile arms-control conversation with two nuclear peers.
- The story is a statement of willingness, not evidence that a meeting has been scheduled or that negotiations have begun.
China’s economy grew 4.5% year over year in the third quarter, beating the cited forecast as demand for high-tech exports helped offset weaker areas of growth.
More
- The National Bureau of Statistics put third-quarter GDP growth at 4.5%, up from 4.3% in the second quarter and above the 4.4% economist consensus cited by the report.
- The article attributes some momentum to global demand for Chinese high-tech exports, but the excerpt does not break down export sectors or their contribution to GDP.
- The result keeps Beijing’s annual growth target within reach, though one quarter’s aggregate figure says little by itself about household demand, property stress or job quality.
- For technology executives, the export contribution is a reminder that China’s growth mix increasingly depends on industrial and technology-linked output, not only domestic consumption.
- The report’s short excerpt does not provide the full GDP release or enough sector data to assess how broad-based the acceleration was.
Operators from hardware, software and enterprise projects say privacy, persistent availability and predictable latency are stronger purchase drivers than raw model size or token savings.
More
- A Shanghai elder-care deployment uses an on-device assistant for conversation, medication reminders and fraud warnings, where sensitive personal data and round-the-clock operation make local processing attractive.
- Enterprise interviewees describe a preference for private deployment when agents must read invoices, orders, financial reports or other core business records.
- Participants say cloud access can be unacceptable for classified or sensitive information because fragments shared over time may reveal a larger whole.
- The commercial hurdle is operational rather than just technical: response time, acoustic noise, reliability, hardware cost and timely model updates determine whether devices remain in use after a demo.
- The panel’s claim that China’s AI will be primarily on-device is an industry forecast, not a demonstrated market-wide outcome.
Washington imposed sanctions on five Chinese entities under a US intellectual-property law, including a Shanghai firm founded by a former Google engineer convicted of trade-secret theft.
More
- The South China Morning Post reports that the State Department sanctioned five entities suspected of stealing American companies’ trade secrets under the Protecting American Intellectual Property Act.
- The named targets include Shanghai Zhisuan Technology, founded by former Google engineer Ding Linwei, venture-capital firm MiraclePlus and three related entities.
- The report says Ding was convicted on seven counts in January; the available excerpt does not detail the sanctions’ exact restrictions or evidence against each entity.
- The move raises the cost of cross-border talent and technology flows and may shape how Chinese AI startups assess access to US-linked research and capital.
The US government ordered civilian agencies to remove DeepSeek products and services built on its model, citing national-security concerns.
More
- The South China Morning Post reports that the Department of Homeland Security directed federal civilian agencies to remove DeepSeek products and related third-party services from their information systems.
- The notice also bars agencies from acquiring DeepSeek products and services, including offerings from affiliated entities.
- The action is a federal procurement and system-use restriction, not a blanket prohibition on private-sector or consumer access.
- For Chinese AI vendors, the order adds another layer of government scrutiny around model provenance and deployment risk in US markets.
Alibaba Cloud describes experiments that move agents from individual developer tools into production workflows, while acknowledging that organization-wide autonomy still lacks a settled operating model.
More
- The account organizes the shift as tool-native work, workflow-native execution and, eventually, organization-native coordination among people and agents.
- Alibaba Cloud frames the agent stack as model, harness and context: the model decides, context preserves relevant state, and the harness governs execution.
- Examples reported by the company include reducing one workflow from two hours to five minutes and shortening another release cycle from 40 days to two; the outlet says these are local results, not aggregate performance.
- Its proposed platform priorities are reliable execution, reusable skills and tools, and organizational controls for agent identity, permissions, observability and human-agent collaboration.
- The company says it is using sandboxes and exploratory metrics rather than treating early experiments as formal performance targets, an attempt to limit local metric gaming.
An autonomous-driving researcher argues that system-level feedback, evaluation and real-world data matter more than cycling through new VLA or world-model designs.
More
- The interview traces the researcher’s work from trajectory prediction to end-to-end driving and then to closed-loop evaluation and interactive simulation.
- His critique of modular driving systems is that information can be lost between perception, prediction and decision modules, while each component may optimize a metric that does not improve the final driving outcome.
- The article describes Think2Drive as using a low-dimensional latent space to reduce the cost of reinforcement-learning interaction in driving simulation.
- The researcher says embodied AI inherits similar problems: models need robust closed-loop evaluation and feedback tied to real task outcomes, not only better open-loop predictions.
- The interview is one researcher’s account of a field-wide shift, not a consensus that end-to-end architectures have solved autonomous driving.
Alibaba’s AliExpress and Tmall say they will pair early-stage investment with domestic and overseas commerce support, moving the platform’s role from selling mature products to helping create brands.
More
- The plan targets more than 100 AI-hardware brands and combines early financing with Tmall launch support and AliExpress marketing, logistics and compliance services.
- The strategic shift is from competing only for listings and traffic to taking an equity position before products have proved themselves in the market.
- For hardware founders, the bundled offer addresses a real gap: engineering teams may know how to build devices but lack overseas customer acquisition, support and fulfillment capabilities.
- For the platforms, investing early may secure future sales and strengthen their position in a category where software updates and post-sale user communities matter.
The Chinese-origin multicluster Kubernetes project reached CNCF graduation and added scheduling improvements aimed at workloads whose GPUs are spread across clusters and clouds.
More
- Karmada extends familiar Kubernetes APIs across multiple clusters, allowing teams to distribute workloads without replacing their existing resource definitions.
- Its control plane uses placement and override policies to express where workloads run and how cluster-specific settings are applied.
- The report says the v1.19 release promotes priority-based scheduling to beta and improves coordination for AI training across clusters.
- The underlying infrastructure problem is fragmented accelerator capacity: a single cluster may not have enough GPUs, while multi-cluster training needs placement, migration and failure handling.
- CNCF graduation signals maturity and governance review, but does not by itself establish that Karmada is the best fit for every production workload.
A Party mouthpiece uses the Long March to argue that the Communist Party needs a central leader whose authority members must uphold, making historical commemoration serve a contemporary discipline message.
More
- The article was published by the ruling Party’s mouthpiece and written by Xie Chuntao, vice-president of the Central Party School, according to the report.
- Its stated lesson is that the Party can unite and carry out orders only when it has a core leadership and members uphold that core’s authority.
- The emphasis is not merely on the Long March as national history; it explicitly links historical success to centralized authority and obedience in the present Party.
- Such commentary is a useful signal of the language the Party is using to reinforce cohesion, though the supplied excerpt does not tie it to a specific personnel or policy change.
Guoyu Xingchen raised nearly 100 million yuan (about $14M) to develop spacecraft servicing robots, moving China’s commercial-space pitch beyond launches toward satellite operations.
More
- The company’s reported funding is nearly 100 million yuan (roughly $14M), and its stated focus spans satellite refueling, repair, assembly and other in-orbit services.
- Its proposed robot combines orbital maneuvering, multi-source AI perception, an onboard decision unit and a dexterous arm for physical interaction with spacecraft.
- The company says its first validation satellite has spent more than a year in orbit and that it has entered a national satellite-internet operator’s supplier list.
- A second demonstration satellite is planned to test a seven-degree-of-freedom arm, a key capability not yet validated in orbit according to the report.
- The strategic bet is that satellite constellations will create demand for maintenance and debris management, but customer willingness to pay and reliable autonomous operations remain unproven.
Fast Retailing’s Western sales have overtaken Greater China, while Uniqlo is shrinking its mainland store base and shifting from rapid expansion toward productivity and localization.
More
- The report says North American and European revenue together surpassed Greater China in Fast Retailing’s latest fiscal year, while mainland China revenue grew only modestly in local currency.
- Uniqlo closed 51 mainland stores in fiscal 2026, taking its count to 851, after a low-tier-city expansion strategy struggled against discounting and online competition.
- The company’s earlier target of 3,000 China stores has receded as it emphasizes closing less productive locations and letting e-commerce serve smaller markets.
- The article links stronger performance in Europe to local product development, a higher share of womenswear and a more locally rooted operating team.
- This is not a China exit: the report says Greater China returned to growth, but the region no longer plays the outsized role it once did in overseas earnings.
Manus’s Cue can manage email but its phone and payment features were inaccessible in the reviewers’ China-based test; the product’s independent-agent vision also follows a major ownership reversal.
More
- Cue gives each agent a separate email, phone number, digital wallet and computer so it can continue working when the user is logged out; the review says email handling was the most complete feature tested.
- The reviewers could not complete phone or payment flows in their environment because they required international accounts; the article cautions that Cue targets overseas users and this test does not represent the full intended-market experience.
- A persistent identity raises a governance problem beyond model capability: who grants an agent authority, how narrowly permissions can be scoped, and who takes responsibility for mistakes.
- The article recounts that China’s National Development and Reform Commission required the reversal of Meta’s proposed acquisition, after which Chinese investors reportedly bought back shares from Meta and the founders regained control.
- Manus has since raised more than $500 million at a reported target valuation near $4 billion, according to the article; its claim that the company may prepare for a listing is explicitly unconfirmed.
- The competitive pressure is rising as large Chinese platforms build their own task agents, making distribution and durable user context harder for an independent agent company to secure.
The PLA publicly denounced a Chinese student accused of wearing the flag at a Russian university, framing the alleged act as whitewashing Japanese militarism and showing how wartime symbols remain politically charged.
More
- The report says the student was accused of wearing the flag in a classroom at a Russian university; the supplied text does not establish the full circumstances or the student's response.
- The PLA statement treats the symbol as an affront to historical memory, rather than as a campus dress-code dispute.
- The episode shows the military using public commentary to reinforce patriotic norms beyond China's borders.
- The flag is also used by Japan's Self-Defense Forces, while in China it is widely associated with Japanese imperial aggression.
MemSilicon says its new funding will support RRAM commercialization and an LPU designed to reduce the memory and data-movement bottlenecks in agent inference.
Also:
Leiphone
More
- The company announced an A-round above 400 million yuan (roughly $56M), with funds earmarked for RRAM commercialization and a new inference processor.
- Its stated target is low latency, long context and high throughput, arguing that repeatedly moving model weights and KV-cache data can dominate inference time.
- The company’s proposed response is system-level coordination of compute, memory and interconnect rather than simply adding more GPU capacity.
- These are product goals, not measured performance claims; the article offers no silicon benchmark or delivered customer system.
Shenzhen urged employers to reduce commuting during the leaders’ meeting, and 3D-printer maker Bambu Lab says local teams may work remotely or on staggered schedules.
More
- The city’s guidance encourages remote work, staggered hours and use of public transit during the APEC meeting week in November.
- Bambu Lab told the outlet that Shenzhen departments may use remote or staggered work arrangements for employee safety and convenience.
- The arrangement shows how a major diplomatic event can reshape ordinary work and mobility across a host city, even without a blanket shutdown order.
- The city’s wording is an encouragement to employers, not a mandatory citywide work-from-home directive.