OpenAI

Sophos cuts threat investigation time by 96% with OpenAI Daybreak

OpenAI

(LLM summarization failed, please manually supplement content later)

Asana cuts model costs 76x in browser tests with GPT-6.1 Sol

OpenAI

Office collaboration vendor Asana carried out performance optimization for browser-side agents, adopting the technical solution of calling GPT-6 Astra (the GPT-6.1 Sol mentioned here) in the Codex module. Actual tests delivered significant cost reduction and efficiency improvement: model usage cost is 76 times lower than the original solution, and running speed is 5 times faster. The company can subsequently rely on this optimization result to provide customers with more capable LLM services.

Anthropic News

Expanding the Cyber Verification Program

Anthropic

This expanded and upgraded Cyber Verification Program (CVP) is selectively open to qualified security professionals. The program provides two core types of resources for compliance practitioners: first, access to advanced cyber technical capabilities, and second, an optimized classifier with a lower false interception rate, which can effectively reduce system blocking interference during legitimate security research and support related work.

2026 Usage Policy update

Anthropic

This is the announcement of the 2026 usage policy update. The official has officially launched the new version of the usage policy, and this announcement is intended to sort out and summarize all adjustments made in the new policy compared to the previous old version, so that entities subject to the policy can quickly understand the key points of the rule changes. The specific complete policy terms are subject to the full version of the new policy text officially released.

Google DeepMind

EmbeddingGemma 2: an open, lightweight multimodal embedding model

Google DeepMind

EmbeddingGemma 2 is an open-source lightweight multimodal embedding model launched by Google, with a far smaller parameter scale than mainstream models of the same type. On benchmark tasks such as image-text retrieval, cross-modal semantic matching, and multimodal classification, its performance matches or even outperforms large-parameter competing products. It supports end-side deployment, can adapt to multimodal retrieval and content understanding needs in low-resource scenarios, and greatly lowers the application threshold of multimodal embedding technology.

Gemini 4 Argon: our next era of frontier intelligence

Google DeepMind

Currently, only the title of the paper Gemini 4 Argon: our next era of frontier intelligence is provided, with no full English abstract attached, so it is not possible to extract, translate, and organize its core methods and conclusions for a summary. Please supplement the full abstract text of this paper, and we will output a concise 120-word English summary as required.

Hugging Face Blog

Impactful scheduling for GPU clusters

Hugging Face

This article addresses the issues of traditional GPU cluster scheduling that fails to accurately match task computing power requirements and does not distinguish business priorities, leading to resource waste and difficulty guaranteeing SLA for high-priority tasks. It proposes the Impactful scheduling strategy that perceives task business value and computing power profiles: pre-model the resource requirements and business weights of tasks, dynamically prioritize scheduling tasks with high adaptability and high impact. Actual tests show that compared with baseline scheduling, GPU utilization increased by 24%, and the time consumption of high-priority tasks was reduced by 32%, balancing resource efficiency and business priorities.

The model that didn’t exist, so you made it yourself

Hugging Face

Currently, only the title information of this paper is obtained, and the core abstract content is not fully pasted, so it is impossible to accurately sort out the core methods, experimental results and core conclusions of the paper. Please supplement the full text of this abstract, and we will extract a concise 120-word English summary highlighting methods and conclusions as required.

Lil’Log

Harness Engineering for Self-Improvement

Lilian Weng

This article sorts out the conceptual context of Recursive Self-Improvement (RSI): In 1965, scholar I.J. Good first proposed the relevant concept, defining that a superintelligent system can surpass all human intellectual activities and independently design better machines to achieve iterative upgrades; in 2008, Eliezer Yudkowsky clarified that its core is the feedback loop where AI relies on existing intelligence to optimize its own cognitive architecture. Current RSI in the AI field can be manifested in two paths: directly rewriting its own weights, or optimizing the training process in a broad sense.

QbitAI

DeepSeek Bot is here! Community version, usable with a plugin on DSH

QbitAI

This is an unofficial DeepSeek-Bot plugin launched by community developers, built based on DeepSeek Harness, and adopts a plugin-based architecture to ensure scalability. It supports customizing the identities of multiple Bots and connected models, comes with a long-term memory function, and allows setting a supervisor Bot to coordinate and schedule other Bots, greatly reducing multi-Agent management costs and token consumption. It will support access to local Agents in the future, and has received more than 250 stars less than 1 day after launch.

Siemens to appear at 2026 Industrial Fair: promote accelerated deployment of physical AI with software-hardware collaborative industrial full-stack capabilities

QbitAI

Siemens will appear at the 26th Industrial Fair 2026, where it will fully display the end-to-end full-process solution for humanoid robots from design to industrial application for the first time, simultaneously launch more than 20 new products, and open 6 types of embodied intelligence factory entry scenarios. Relying on its software-hardware collaborative industrial full-stack capabilities, integrating technology and industry accumulation, it promotes the deployment of physical AI to the production frontline, helps customers improve efficiency and resilience, and drives intelligent growth through the integration of digital and physical worlds.

24 days! Jev’s valuation rose to $7.5 billion, probably Silicon Valley’s best underdog story of the year

QbitAI

Founded in 2024, San Francisco startup TypeSafe AI previously stayed under the radar in Silicon Valley. Its launched AI model Jev cuts out the generation function of traditional large models and only focuses on binary decision-making, with extremely low cost and token pricing far below the industry level. Only 24 days after launch, the company completed an $870 million Series A financing, with a post-money valuation of $7.5 billion, receiving heavy investment from top capital firms such as a16z and Sequoia.