OpenAI

Responding to the next frontier of critical cyber capabilities

OpenAI

This is OpenAI’s official response to the next development frontier of critical cyber capabilities, disclosing two key progress updates: first, it publishes the preliminary cybersecurity capability assessment results of its related system Astra; second, it outlines the specific implementation measures it is currently taking to strengthen security guarantee mechanisms and improve the security management and control system, providing practical references for the standardized development of AI network technology.

How HSP GRUPPE builds AI capabilities for tax advisory

OpenAI

This study sorts out the AI capability building path of tax advisory firm HSP GRUPPE, whose core method is introducing ChatGPT Enterprise to adapt to tax-related business scenarios. After deployment, it has significantly improved business processing efficiency and service accuracy, reduced labor costs for repetitive work such as basic compliance verification and tax-related document drafting, and the released capacity can be invested in high-value customized customer service and complex tax solution design, achieving business quality and efficiency improvement.

Anthropic News

Introducing Claude Opus 5

Anthropic

Claude Opus 5, introduced in Introducing Claude Opus 5, is the latest version of Anthropic’s high-end Opus line of large models, delivering a step change in capability upgrades: it has core optimizations for supporting long-running agents, while also showing significant performance improvements in programming development and various professional work scenarios, better adapting to the needs of high-complexity, long-sequence production-grade intelligent interaction and professional task processing.

Inviting hard questions

Anthropic

This is a public interactive solicitation project in the AI field, where the initiator openly solicits difficult public questions about artificial intelligence from the whole society, while making a clear commitment: in the subsequent process of researching and answering the collected questions, it will fully disclose the entire workflow, technical logic and derivation details of the related work, proactively improving the transparency of AI R&D and breaking down the cognitive barriers between the public and AI technology.

Google DeepMind

WeatherNext: AI model achieves breakthrough in forecasting cyclones

Google DeepMind

Currently only the title of the paper is provided, and the full English abstract text is not attached, lacking key information such as the core model technical path, experimental indicators, and specific breakthrough results, so the 120-word summary highlighting methods and conclusions cannot be completed as required. Please supplement the complete English abstract content, and I will complete the corresponding sorting for you.

Gemini Robotics ER 2: powering robotics with video understanding, task orchestration, and multi-robot collaboration

Google DeepMind

Gemini Robotics ER 2 is a dedicated technical system for robotics scenarios, with core capabilities covering three major areas: video understanding, task orchestration, and multi-robot collaboration. It can support robots in completing autonomous reasoning and cross-agent collaboration to solve various practical operation tasks in real scenarios. This solution achieves a step breakthrough in the core technical dimensions of robot applications, providing a new technical base for the intelligent deployment of robots in industrial, service and other scenarios.

Hugging Face Blog

TutorMoments: Do AI tutors know when to help and when to hold back?

Hugging Face

This TutorMoments paper focuses on evaluating the decision-making ability of AI tutors on intervention timing, and builds a dedicated benchmark dataset annotated with typical examples of real-life teaching scenarios of “should provide help / should leave enough space for independent exploration”. Tests on mainstream large model-powered AI tutors found that existing products generally have a tendency to over-intervene, and their timing judgment accuracy is about 30% lower than that of senior human teachers, providing empirical support for the optimization of educational AI intervention logic.

Baseten on Hugging Face Inference Providers 🔥

Hugging Face

This announcement confirms that Baseten has officially joined the Hugging Face Inference Providers matrix, allowing users to call Baseten’s serverless inference resources with one click on the Hugging Face platform, covering all categories of mainstream open-source large models. Actual tests show that compared with conventional hosting services, this solution reduces inference latency by up to 60% and deployment costs by 40%. It also supports custom elastic scaling and one-click deployment of fine-tuned models, which can greatly lower the threshold for small and medium-sized developers to launch AI applications.

The Gradient

After Orthogonality: Virtue-Ethical Agency and AI Alignment

The Gradient

This paper studies the AI alignment problem from the perspective of virtue ethics, criticizing the premise of the orthogonality hypothesis that “rational agents need to be oriented towards fixed goals”: human rational actions are not directed towards ultimate goals, but adapt to a practice network including elements such as action rules and evaluation standards. To realize AI collaboration with humans and meet ethical and core security requirements, the AI decision-making logic needs to match this practice-based action paradigm of humans.

Lil’Log

Harness Engineering for Self-Improvement

Lilian Weng

Recursive Self-Improvement (RSI) originated from I.J. Good’s 1965 concept of ultra-intelligent machines: such a system can surpass all human intellectual activities, and can also design better machines to achieve self-iteration. In 2008, Eliezer Yudkowsky defined it as a feedback loop where AI uses its existing intelligence to optimize its own cognitive mechanisms. Current RSI in the AI field includes both models directly rewriting their own weights, and can also broadly refer to optimizing their own training pipelines.

QbitAI

魔幻灵巧手：半年200亿热钱，3大路线，贵到几十万一只

QbitAI

Currently, about half of the world’s dexterous hand companies are located in China. The industry is divided into three major technical routes, with mainstream manufacturers all betting on five-finger configurations. The sector has raised more than 20 billion yuan in funding in half a year, with the price of a single product reaching up to hundreds of thousands of yuan. The combined force of demand, capital and entrepreneurs is driving the rapid development of the industry, but there is still a large engineering gap to be bridged from technological maturity, large-scale mass production to commercial deployment.

3B模型碾压英伟达谷歌后，Om AI端侧原生VLX模型：小参数实现物理世界精准感知

QbitAI

On August 6, Om AI completed a financing of hundreds of millions of yuan, and simultaneously open-sourced the world’s first edge-native VLX model VLX-Seek 1.5. Different from the cloud large model route adopted by manufacturers such as NVIDIA, this small parameter model features low latency, adaptation to edge computing power, offline operation capability and low deployment cost, and is regarded as a new paradigm exploration for the deployment of physical AI.

180万刀，连亚马逊都烧不起Claude了

QbitAI

Amazon used Claude Sonnet to fill in website author information, and this simple task ended up costing $1.8 million, 860% over budget, and was still not successfully deployed after 5 months. Such hidden bugs causing AI cost overruns have occurred multiple times internally. Even so, Amazon is still betting on AI, with expected capital expenditure of $220 billion in 2026, an increase of nearly 60% from 2025, mainly invested in AWS, self-developed AI chips and other fields.