OpenAI

New usage analytics and updated spend controls for enterprises

OpenAI

OpenAI has recently rolled out two new capabilities for ChatGPT Enterprise: a brand-new usage behavior data analysis module, and iteratively upgraded spend control tools. These two features precisely meet the cost management needs of enterprise AI applications, helping enterprises clearly monitor AI usage expenses, reduce uncertainty in large-scale AI implementation, and support enterprises in advancing their AI business layout more steadily.

Improving health intelligence in ChatGPT

OpenAI

This study focuses on upgrading ChatGPT’s intelligent response capabilities for health-related queries, with optimization primarily implemented based on GPT-5.5 Instant: technically, it enhances reasoning ability, optimizes context adaptation, and improves expression clarity, while introducing an evaluation and calibration mechanism with physician participation. Ultimately, it greatly improves the professionalism and reliability of ChatGPT’s answers in health and wellness scenarios, delivering accurate health-related responses that better fit user needs.

Anthropic News

Statement on the US government directive to suspend access to Fable 5 and Mythos 5

Anthropic

This statement addresses the latest export control directive from the US government: US authorities have issued control requirements to fully suspend access to the two products Fable 5 and Mythos 5 for all foreign nationals, whether they are located inside or outside the United States. This measure is a restriction on the export of cutting-edge technologies, which will directly affect cross-border R&D cooperation and overseas commercial deployment of related technologies, highlighting the protectionist tendency in the US high-tech sector.

Introducing Claude Corps

Anthropic

The newly launched Claude Corps is a nationwide talent training and funding program, established specifically for early-career practitioners who are passionate about promoting the inclusive value of AI technology. Its core purpose is to support these young talents to leverage their abilities, effectively deliver the various benefits and values generated by AI development to grassroots communities across the United States, and fill the gap in the last mile of AI inclusion implementation.

Google DeepMind

Unlocking UK house-building with AI-accelerated planning

Google DeepMind

The UK government has collaborated with Google DeepMind to develop an AI-driven housing planning approval prototype system, targeting the current pain point of low approval efficiency in the UK’s housing planning process that restricts housing supply. It relies on AI technology to speed up the planning and decision-making process related to housing construction. Once deployed, it is expected to significantly shorten the planning review cycle and remove pre-construction bottlenecks for housing development, marking a new practice of AI empowering government approval and implementing in the people’s livelihood field.

Securing the future of AI agents

Google DeepMind

This paper focusing on the future security of AI agents proposes an AI control roadmap as the security construction framework. Its core idea is to combine traditional mature security protection mechanisms with real-time dynamic monitoring capabilities to achieve comprehensive protection of internal systems associated with AI agents. This solution takes into account the reusability of existing security systems and the demand for dynamic risk response in AI scenarios, providing a clear construction reference path for the secure development and large-scale deployment of AI agents.

Hugging Face Blog

MosaicLeaks: Can your research agent keep a secret?

Hugging Face

This paper titled MosaicLeaks: Can your research agent keep a secret? focuses on the privacy risks of autonomous research agents, and proposes the MosaicLeaks data leakage test method: split sensitive information into discrete fragments and mix them into reference materials for research tasks, inducing the agent to spontaneously splice, restore and leak the information during task execution. Tests show that the leakage rate of mainstream research agents is as high as over 85%, and existing conventional security protections are almost ineffective, so targeted optimization of confidentiality mechanisms is urgently needed.

Beyond LoRA: Can you beat the most popular fine-tuning technique?

Hugging Face

At present, only the title of this paper is provided, and the full text of the abstract is not attached. Please upload the complete English original text of the abstract, and I will refine and translate it into a Chinese summary of about 120 words as required, highlighting the core methods and research conclusions without redundant repetition of the original content.

The Gradient

After Orthogonality: Virtue-Ethical Agency and AI Alignment

The Gradient

This AI alignment study exploring extensions to the orthogonality thesis takes virtue ethics as the perspective, refutes the presupposition that “rational agents are necessarily guided by fixed goals”, points out that human rationality adjusts actions based on a practical network including elements such as behavior and evaluation standards, and proposes that for AI to adapt and collaborate with humans, its decision-making logic needs to match the human practical reasoning model. This path is related to both AI ethical alignment and core security.

Qbitai

英伟达开始搞机器人自己研究机器人那套了…

Qbitai

NVIDIA, together with CMU and UC Berkeley, has launched ENPIRE, an embodied intelligence autonomous research framework, equipped with 8 coding agents each controlling a dual-arm robot, which can independently complete the full process of literature retrieval, algorithm iteration, automatic reset of real scenarios, and experimental analysis without human intervention. In the 4mm pin insertion task, it increased the success rate from 0 to 99% in only 3 hours, and can conduct uninterrupted autonomous research 24/7, greatly reducing the labor cost of scientific research.

陶哲轩12年前的预言，现在AI帮他兑现了

Qbitai

Terence Tao predicted 12 years ago that mathematical papers would be written in computer-recognizable formal languages in the future, which was regarded as a fantasy at the time. Now breakthroughs in AI in the field of mathematics are accelerating: the “AI + human collaboration” project led by him has solved most of the 22 million mathematical problems in only 48 hours, and a large number of mathematical proofs can now be automatically verified formally, fulfilling his own prediction from years ago.

撸猫撸出SOTA！3个00后2个月，造出史上最快流式音视频社交模型

Qbitai

Three post-2000s developers from domestic startup team Catnip spent 2 months building MaineCoon, a streaming audio-video social model with SOTA performance. This model is the first to achieve synchronized audio-visual streaming generation of over 30 minutes. Its 22B parameter version reaches an inference speed of 47.5 FPS on a single H100, ranking first in the industry, with first frame generation taking less than 3 seconds, and inference cost only 1/560 to 1/2000 of similar products. It has already received active cooperation invitations from leading manufacturers.