Blog

  • OpenAI Launches Embedded Safety Evaluations with Four External Organizations

    OpenAI announced the signing of first-phase embedded evaluation agreements with Apollo Research, Goodfire, METR, and Transluce. The four organizations will each participate in model alignment, interpretability, reinforcement learning runs, and safety evaluations. METR will inspect OpenAI’s largest reinforcement learning run and gain access to relevant checkpoints, rollouts, and monitoring records, but OpenAI retains the right to redact confidential information. As a result, the evaluation findings still do not constitute independent verification.

    Source:Tech Times · Read Original

  • Centered Ablation Explains Misalignment and Collapse in Final-Layer Steering Ablation and Provides a Closed-Form Correction

    The study introduces Centered Ablation (CA), which uses closed-form projection to remove the common-mode component from behavioral directions, addressing cases where final-layer steering ablation has limited effect or causes text collapse. Across eight models, CA increases attack success rates by 29 and 46 percentage points when standard methods have limited effectiveness, and reduces Gemma-2-9B’s collapse rate from 50% to 1%. On models with high overlap, it also preserves 12 to 40 percentage points more normal responses.

    Source:Hugging Face 社区论文精选 · Read the original

  • Amazon Plans to Invest $7 Billion in Renewable Energy Projects in India by 2030

    Amazon plans to invest $7 billion in renewable energy projects in India by 2030 and has added four utility-scale wind and solar projects with a combined capacity of 934 MW. This investment builds on the $4 billion in previously contracted projects. Once operational, the projects will have a combined capacity of more than 2 GW, enough to meet the electricity needs of more than 3.5 million Indian households for one year. Amazon has also set a global goal to reduce the carbon intensity of electricity used in its operations by 65% from the 2019 baseline by 2030.

    Source:About Amazon · Read original

  • China’s Basic Medical Insurance to Cover CAR-T Cell Therapy relma-cel

    China’s National Healthcare Security Administration announced that relma-cel, a CAR-T cell therapy developed by JW Therapeutics, will be included in the basic medical insurance system starting Jan. 1, marking the first time this type of therapy has entered China’s national medical insurance coverage. The therapy has been approved for three types of lymphoma, with a price of 1.29 million yuan per dose; as of the end of 2025, China’s basic medical insurance program had 1.33 billion insured participants. Before inclusion in the medical insurance system, patients typically had to pay the full cost of medicines not listed in the medical insurance catalog.

    Source:Fresh voices from today's China · Read Original

  • Microsoft’s Widespread Outage Should Make Every Business Wary of Cloud Concentration Risk

    Outages affecting Microsoft 365, Outlook, Teams, Exchange Online, and Copilot across more than 70 countries show that businesses need to reassess the risks of cloud concentration and cascading architectural failures.

    Source:InfoWorld · Read Original

  • US Tightens AI Chip Export Controls to Close Remote Computing Loophole at Third-Party Data Centers

    The US Department of Commerce announced tighter export controls on AI chips, requiring US authorization to remotely provide advanced chip computing power to companies in China, Macau, and countries subject to US weapons embargoes. Exporting advanced chips to Singapore, Malaysia, Thailand, Indonesia, and Brazil also requires a license to reduce the risk of chip diversion. Nvidia shares fell 4% on Monday, with cumulative losses exceeding one-quarter over the past month; after Jensen Huang stepped down as CEO, Colette Kress is serving as interim CEO, and the company will outline its future sales, Asian operations, and AI investment strategy in its earnings report.

    Source:citizen.digital · Read Original

  • NATO Uses OpenAI Cybersecurity Models to Protect Alliance Networks

    NATO is using OpenAI’s cybersecurity models to protect allied networks through the company’s Daybreak program, consuming up to 300 billion tokens per month for software reviews, suspicious activity investigations, vulnerability validation, and the development and testing of patches. OpenAI also provides models to ENISA and the national cybersecurity agencies of France, Germany, and Poland. ENISA has used them to identify four vulnerabilities, while Poland’s CERT Polska has discovered six vulnerabilities in MikroTik RouterOS, with related fixes already released.

    Source:UK Defence Journal · Read Original

  • Taiwan Stocks Fall 6.03%; TSMC and Nvidia Events Dominate Asian Chip Market

    The Taiwan Weighted Index fell 6.03% to 41,040.08 points on Nov. 16, with TSMC down 5.74%, dragging the index by about 1,000 points. South Korea’s Kospi fell 3.59%, while SK Hynix dropped 7.89%. Nvidia will release its earnings on Wednesday, and Taiwan index futures will be cash-settled against the average index in the final 30 minutes of trading. The market is closely watching the impact of Nvidia’s earnings results and changes in export controls on chip stocks and settlement.

    Source:Tech Times · Read Original

  • Chinese Institutions Use U.S. AI Model Outputs to Train Domestic Models Amid Chip Restrictions

    After reviewing more than 300 Chinese academic papers, patents, corporate disclosures, and government documents, Reuters reported that Chinese institutions are using the outputs of leading U.S. AI models to train domestic competitors, with the practice having moved beyond academic experiments into industrial applications. Researchers selected approximately 5,000 answers from OpenAI o1 and translated them using GPT-4o mini to create a bilingual dataset of 10,750 training examples, which was then used to further train Alibaba’s Qwen2.5-72B.

    Source:devdiscourse.com · Read Original

  • US Introduces New Rules Restricting China’s Remote Access to Advanced Computing Chips and Servers

    The Trump administration released and immediately implemented new rules restricting remote access to US advanced computing chips and servers through overseas data centers, preventing third-party cloud computing arrangements from circumventing export controls. Overseas data centers serving China, Russia, Iran, and Cuba must identify end users, while selling relevant chips to Indonesia, Malaysia, Singapore, Thailand, or Brazil requires additional licensing requirements. China’s Ministry of Foreign Affairs accused the United States of interfering with trade and technological cooperation and said it would take necessary measures to safeguard its own interests.

    Source:WTOP News · Read original