Daily economy, technology & research · Two language editions
Today’s useful developments, in about ten minutes
Choose a complete Chinese or English edition covering the same stories.
English edition 11:04
Audio could not load. Try the download link or retry later.
AI-generated voices · Releases and reporting below · Play when ready
Today’s briefing
Welcome to the daily economy, technology and research briefing for November 6, 2026. Today we cover economic developments in China and around the world, alongside useful technology and research updates. Each story comes with an attributable source. We explain the reporting period and the limits of the evidence, and keep any practical suggestions separate from the news.
1. Shandong’s Yellow River Cities Combine Smart Manufacturing with Green Energy
Xinhua Finance reported smart factories, industrial AI, vehicle production and offshore wind equipment across several Yellow River cities in Shandong. Provincial statistics showed a 7.4% year-on-year rise in industrial value added above designated size in the first three quarters, while equipment manufacturing grew 12%.
China economy · 新华社财经 · 2026-11-06 · Original report
Read the full narration
Xinhua Finance reported from several cities along Shandong’s Yellow River section, showing manufacturing upgrades and green-energy development progressing together. In the Jinan New and Old Kinetic Energy Conversion Starting Zone, an Aiko solar factory uses energy-saving equipment and is building an integrated source-grid-load-storage system to increase its green-power share. The report also visited older industrial centers. In Gaoqing County, six upgraded robotic arms at Zibo Bohai Piston reportedly changed a manual operation into equipment-assisted production. Company representative Zhao Yongquan said labor efficiency improved by more than 200%. That is a company statement, not an independently measured industry average. In Zouping, Shandong Hongao deputy general manager Li Shuguang said its lightweight body is 40% lighter than a conventional steel body and could increase an electric vehicle’s range by 3% to 5%. Those performance figures also come from the company. For wider context, Shandong’s provincial statistics bureau reported that industrial value added above designated size grew 7.4% year on year in the first three quarters, with equipment manufacturing up 12%. The main limitation is that this is a reported selection of enterprise cases, not a comprehensive industrial census. Local examples therefore cannot establish average productivity or energy performance across Shandong.2. ECB Publishes Consolidated EU Banking Data for End-June 2026
The ECB reported that EU-headquartered credit institutions’ total assets rose 5.18% to €34.99 trillion in June 2026 from €33.27 trillion a year earlier. Their aggregate non-performing loan ratio increased 0.03 percentage points year on year to 1.99%.
World economy · European Central Bank · 2026-11-06 · Original release
Read the full narration
The European Central Bank published consolidated banking data for the EU banking system as of end-June 2026, compiled on a group-consolidated basis. It reported that the aggregate total assets of EU-headquartered credit institutions increased by 5.18%, from €33.27 trillion in June 2025 to €34.99 trillion in June 2026. Over the same period, the aggregate non-performing loan ratio rose by 0.03 percentage points year on year to 1.99%. The aggregate return on equity was 5.31%, while the Common Equity Tier 1 ratio stood at 16.43%. The accounting details matter. The ECB defines the non-performing loan ratio as non-performing loans divided by total loans. Its return-on-equity figure uses total profit or loss for the first six months of 2026 divided by total equity and is not annualized. It therefore should not be equated with a full-year return. Most reporters generally apply International Financial Reporting Standards and the European Banking Authority’s supervisory reporting standards, although some smaller reporters may use national accounting standards. A few historical figures were also revised in this release. These figures describe several separate dimensions of banking health, not one overall score. They are EU-level aggregates, so they cannot establish the condition of an individual bank or member state. The non-annualized profitability measure is also not directly comparable with annual figures.3. NVIDIA Adds Four Open-Weight Cosmos 3 Simulation Checkpoints
NVIDIA published four Cosmos 3 Nano post-training checkpoints for driving-data generation, streaming driving simulation, simulator-to-image conversion and bimanual robot video prediction via Hugging Face. NVIDIA said each checkpoint has 16 billion trainable parameters.
Technology & research · Hugging Face · 2026-11-05 · Primary source
Read the full narration
NVIDIA announced four new open-weight checkpoints for its Cosmos world-model development platform through Hugging Face. The company said all four are post-trained examples of Cosmos 3 Nano, have 16 billion trainable parameters, and use OpenMDW 1.1, which permits commercial and non-commercial use under its terms. Each checkpoint targets a different input-output relationship. Transfer-Auto generates offline synthetic driving data, synchronizing video across as many as 11 camera views and supporting joint RGB and LiDAR generation. Sim-Auto is its causally distilled streaming variant, producing successive chunks of synchronized camera video and LiDAR from generated history and updated driving controls. A closed-loop application would additionally require a simulation runtime and a World Scenario Map renderer to update those controls. Sim-Transfer takes spatial controls from a physics simulator and a text prompt to produce color observations, while motion, contacts and physical state remain with the simulator. Sim-Bimanual predicts the next observation from a scene and robot or camera actions and supports interactive two-arm simulation for YAM and AgiBot embodiments. The limitations are substantial. NVIDIA requires Linux, an NVIDIA GPU, Docker and NVIDIA Container Toolkit, and says the published vLLM-Omni setup was tested on one GB200, which is not a minimum-memory specification. It also warns that errors in self-generated history can compound. A successful short run therefore demonstrates execution, not convergence or reliable policy evaluation.4. Gemini Omni 1.2 Flash Adds Longer Video and Conversational Editing
Google said Gemini Omni 1.2 Flash is generally available through the Gemini API and Google Cloud, with native video generation up to 15 seconds and multi-turn conversational editing. It said the 720p price is $0.07 per second, 30% below version 1.1.
Technology & research · Google AI · 2026-11-05 · Primary source
Read the full narration
Google AI announced Gemini Omni 1.2 Flash on November 5 and said the model is generally available through the Gemini API and Google Cloud, with access also offered across products including Google Vids and Gemini for business. According to Google, the new version can generate and edit video natively for as long as 15 seconds, up from 10 seconds in Omni 1.1, and allows duration to be specified in one-second increments from three to 15 seconds. The company also described multi-turn conversational editing. Users can revise the same clip to swap a product or background, change lighting or weather, or switch the aspect ratio rather than start over. Google said the model can combine character, product or style images, motion clips of up to three seconds and audio cues, with improved adherence to details such as clothing, packaging typography and brand colors. These are product claims rather than results from an independent evaluation. Pricing is another practical change. Google stated that 720p generation costs $0.07 per second, 30% less than Omni 1.1. The limitation is that Google describes generative AI as experimental and specifically says rapid sports action, fast script changes and complex physics still need improvement. Vendor examples therefore cannot substitute for repeated tests under local teaching conditions, and they do not guarantee consistent identity, text, continuity or physical behavior in every generation.5. Agile Deliberation Makes Subjective Image Labeling Interactive
Google Research presented Agile Deliberation, which helps users define subjective image categories through concept scoping, borderline-example feedback and automatic prompt refinement. Its evaluation included 18 live sessions of 1.5 hours each, and the underlying code was open-sourced.
Technology & research · Google Research · 2026-11-04 · Primary source
Read the full narration
Google Research introduced Agile Deliberation, an interactive framework for building image classifiers around subjective categories whose boundaries may change with context or operational policy. Examples in the research include “healthy food” and “inspiring artwork.” Rather than requiring people to act as static labelers, the framework treats category definition as an iterative conversation. The workflow has two stages. In concept scoping, the system helps a user divide a broad target into positive subcategories and explicit exclusions. In concept iteration, vision-language models retrieve semantically borderline images that challenge the current written definition. A user judges these examples, identifies blind spots and supplies natural-language feedback. That feedback enters an automatic prompt-optimization loop so the classifier’s boundaries change alongside the clarified intention. Google Research said the underlying code and algorithms were open-sourced on GitHub, with a paper and interactive Colab or Jupyter demonstration also linked. The evaluation included 18 live user sessions, each lasting 1.5 hours, with participants encountering the framework and alternative baselines in randomized order. The two tasks were Healthy Food and Pay-to-Play, the latter involving misleading clickbait promising monetary or digital rewards. Google Research reported higher labeling consistency, classifiers that outperformed zero-shot baselines, and lower reported cognitive effort and stress than with manual prompt engineering. These are the authors’ findings from the specified study. The evidence comes from two tasks and 18 live sessions, so it does not establish universal performance.That concludes today’s briefing. Before adopting any of these tools or drawing wider conclusions from the figures, check the relevant methodology, public code or documentation, and results from tests conducted under your own conditions. Vendor demonstrations do not replace independent evaluation.
Prepared from the listed releases and reporting with AI-assisted editing and automated checks. Audio is AI-generated. Economic coverage is a factual overview, not investment advice.