OpenAI

Responding to the next frontier of critical cyber capabilities

OpenAI

This material released by OpenAI responds to the demand for the new frontier of critical cyber capability development, and discloses two core points: first, the preliminary results of a special cybersecurity assessment for its Astra system; second, the specific implementation measures the company is currently taking to build a solid security protection barrier and improve the full-process security management and control mechanism, providing a reference for proactive prevention and control of cybersecurity risks in AI scenarios.

How HSP GRUPPE builds AI capabilities for tax advisory

OpenAI

This article introduces the AI capability building practice of professional tax consulting firm HSP GRUPPE: the core method is to introduce ChatGPT Enterprise as a business empowerment tool. After implementation, it effectively improves the processing efficiency of tax business, optimizes the quality of professional output, and at the same time releases internal human capacity, which can be more tilted to high-value-added customized customer service links, providing a reference case for professional service institutions to implement generative AI.

Anthropic News

Introducing Claude Opus 5

Anthropic

The newly released Claude Opus 5 is a stepwise iterative version of the Opus-tier large model. The core upgrades focus on two major directions: first, the support capability for long-running agents has achieved a leapfrog improvement; second, the performance of coding task processing and professional work scenarios in various fields has been significantly optimized, which can better meet the needs of complex agent development and professional production.

Inviting hard questions

Anthropic

This AI research team has launched a public interaction project: it openly solicits the most difficult and confusing questions from the public in the field of artificial intelligence from the whole society, and at the same time makes a clear commitment: when responding to and answering these collected questions, the complete process of research and demonstration work will be disclosed to the public to ensure full transparency of the answering process and actively accept public supervision.

Google DeepMind

WeatherNext: AI model achieves breakthrough in forecasting cyclones

Google DeepMind

This research launches the AI weather forecasting model WeatherNext, addressing the pain points of traditional cyclone forecasting that relies on numerical simulation, has high computing power costs, and insufficient prediction accuracy in extreme scenarios. It integrates multi-source meteorological observation data to train a spatiotemporal sequence prediction framework, which can output results in seconds. Compared with traditional models, it can accurately predict cyclone generation, path and intensity 7 days in advance, with an average error reduced by more than 30%, achieving a key breakthrough in the field of cyclone forecasting.

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

Google DeepMind

The newly released Gemini Robotics ER 2 is a special technical system for robot scenarios, which can provide robots with capability support for environmental reasoning, group collaboration, and real-scenario task solving. The system has achieved stepwise breakthroughs in three core technical directions adapted to robots: video understanding, task tool orchestration, and multi-robot collaboration, which can effectively empower the implementation of various types of robots.

Hugging Face Blog

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

Hugging Face

Currently you have only provided the title of this paper, without attaching the corresponding English abstract content, so it is impossible to complete translation and refinement, and organize it into a summary of about 120 words highlighting the method and conclusion. Please supplement the complete abstract text, and I will sort out the core information for you as required.

Baseten on Hugging Face Inference Providers 🔥

Hugging Face

You have not pasted the corresponding abstract content yet. If it is the official announcement of Baseten accessing Hugging Face Inference Providers, you can refer to the following summary: Baseten has officially become an official Hugging Face inference service provider. Developers can call its hosted open-source models such as large language and multimodal models with one click on the Hugging Face platform, supporting low-latency high-concurrency deployment and custom scaling. Compared with self-built deployment, the cost is reduced by about 40%, which greatly lowers the threshold for AI model implementation.

The Gradient

After Orthogonality: Virtue-Ethical Agency and AI Alignment

The Gradient

This article addresses the AI alignment problem from the perspective of virtue ethics, challenges the implicit premise of the orthogonality hypothesis that “rational agents need to act around fixed ultimate goals”, and points out that the core of human rationality is to anchor behavior to a shared practice network that includes elements such as action norms and evaluation standards. It argues that AI decision-making logic needs to match the practical thinking paradigm of humans, rather than being tied to fixed goals, in order to balance ethical alignment and core security requirements.

Lil’Log

Harness Engineering for Self-Improvement

Lilian Weng

The concept of Recursive Self-Improvement (RSI) was first proposed by I. J. Good in 1965, referring to “superintelligent machines” that can surpass all human intellectual activities and design better machines by themselves to achieve iterative upgrades. In 2008, Eliezer Yudkowsky defined it as a specific feedback loop in which AI optimizes its own cognitive architecture relying on existing intelligence. In the context of contemporary AI, such loops include both the model directly rewriting its own weights and, in a broad sense, training process optimization.

QbitAI

Sam Altman’s ChatGPT parenting method has caused a huge uproar

QbitAI

OpenAI CEO Sam Altman recently shared his ChatGPT parenting plan: integrate family schedules and children’s interests to generate exclusive podcasts for the commute to school, saving parents the trouble of finding common topics. This idea was widely mocked by netizens and criticized as anti-human. The reply from the creator of Gravity Falls saying “just chat with your kid directly” got 18 times more likes than the original post. Netizens resent AI taking over parent-child communication, and coupled with Altman’s previous statement that not having time to accompany his children is his biggest regret, the contradiction in his public image has intensified the controversy.

China’s NeoLab moment: EverMind delivers the first full-stack self-evolution solution with 3 papers

QbitAI

In 2026, when overseas NeoLabs are still hyping concept financing, the AI team EverMind incubated by Shanda has released three papers in a row, delivering the first full-stack solution in the field of self-evolving AI. The team inherits the long-term research gene of Shanda Innovation Institute, does not take the shortcut of traffic monetization, focuses on overcoming the problem of AI long-term memory and exploring the underlying self-evolution path, which is a landmark achievement of China’s local NeoLab model.

Jeff Dean’s startup BP exposed, Yang Zhilin is also on it! Silicon Valley VCs are scrambling to invest

QbitAI

The minimalist business plan of Discovery Loop, a new AI company founded by Jeff Dean, Yang Zhilin and three others, has been exposed. It skips the product, market, and profitability demonstrations required for a regular BP, and only lists the past achievements of the four people - which almost cover Google’s core products, underlying infrastructure, and the full range of AI research and applications. With the top-tier team background, Silicon Valley VCs are scrambling to invest, and the lead investor is proud to get the investment qualification. Netizens joked that their BP only needs to write “Google me” to be sufficient.