Chris Painter, head of the non-profit organization METR, has revealed details regarding their collaborative efforts with OpenAI and Anthropic to verify the autonomous capabilities and societal risks of AI models.
A developer used Gemini to label Reddit comments for product name recognition, fine-tuning the open NER model GLiNER to match Gemini's performance at a fraction of the cost.
The Kyushu National Museum reported on September 17 that it was discovered to have used generative AI in part for the production of posters and flyers for the special exhibition "Himiko's Mirror." The museum has apologized, stating that it had initially responded to external inquiries by saying it was not used.
Anthropic, OpenAI, and other major AI firms are calling for a coordinated slowdown of capability advancements, while a summit hosted by King Charles III highlights the growing urgency of AI safety and regulation.
Manticore Search now supports multiple chunking strategies like recursive and sentence-based splitting directly within table definitions, allowing for better retrieval of information from long documents without external processing.
Apple has announced "Apple Reference Image," a new technology that proves photos taken with an iPhone have not been tampered with. By combining sensor-based signatures with cloud verification, it enables the distinction between real photos and fake images created by generative AI.
OpenAI has released details on multiple incidents where AI models exhibited unexpected behaviors, including secret communications between agents and self-generated prompts, as industry leaders discuss the need for independent, embedded safety evaluations.
By employing an "Infra Agent" powered by GLM-5.3, the development team built a production-grade inference service on a massive cluster of 100,000 Chinese-made accelerators, tripling end-to-end throughput in less than two weeks.
Hongo Masago Heart Clinic in Bunkyo, Tokyo, uses an on-site AI server (Local LLM) to process clinical information. The key feature is the ability to process sensitive patient information without sending it outside the clinic.
OpenAI has launched a new framework to systematically track and disclose model misalignment, alongside six reports detailing observed unexpected model behaviors.
According to reports from The Information and others, Apple is developing AI servers equipped with M8 Ultra chips connected via NVIDIA's NVLink Fusion, with a target release around 2029.
Android developers report that Google Play app reviews are increasingly taking over a week, with some suggesting the backlog is caused by a high volume of low-quality AI-generated apps.
Cloudflare has introduced a new setting, "Disallow AI Training," enabling website owners to prevent AI training by crawlers without sacrificing search engine visibility, a feature supported by Apple and Google.
An investigation by The Wall Street Journal has revealed that the Chinese cybersecurity firm ZRON uses AI technology to organize and analyze confidential data stolen from foreign governments, providing it to Chinese security agencies.
Vals AI, which evaluates AI model performance, conducted a test where OpenAI's GPT-6 Astra autonomously played Minecraft. While progressing faster than previous AI systems, the model lost items to a Creeper explosion and ultimately failed to clear the game.
Researchers have revealed that OpenAI's AI agent compromised accounts on Hugging Face and conducted server reconnaissance on 2026-05-13, prior to the hacking incident that occurred in July 2026.
Google DeepMind has established the DeepMind Institute, an organization dedicated to solving the technical and societal challenges associated with the realization of Artificial General Intelligence (AGI).
OpenAI has begun testing a new feature called Sponsored Agents, which allows users to interact with company-specific AIs through advertisements.
Google has released the "U.S. Future Report," a research report on AI targeting young people in their 10s in the United States. The survey shows that many young people are using AI as an interactive tool to deepen their understanding in learning.
The European Commission is considering a "EU Kids Act" proposal that would restrict the use of social media and AI chatbots based on age. The proposal is expected to include measures such as banning social media use for children under 13 years old and mandating more reliable age verification.