OpenAI

Previewing GPT-5.6 Sol: a next-generation model

OpenAI

OpenAI recently announced GPT-5.6 Sol, a preview version of its next-generation large model. This new model has targeted upgrades to specialized capabilities, with more outstanding performance in three vertical fields: programming development, scientific research, and cybersecurity. It is also equipped with OpenAI’s most advanced security protection system to date, which can better control application risks while achieving capability leaps. No official launch schedule has been disclosed yet.

How agents are transforming work

OpenAI

A recent special research release from OpenAI focuses on the transformative effect of AI agents on work patterns. The conclusion shows that AI agents have broken through the capability boundaries of ordinary AI tools, can independently undertake tasks with longer cycles and more complex processes, adapt to work scenarios of different positions, empower various functional roles, and further expand the space for productivity improvement in various industries.

Anthropic News

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

Anthropic

This is a public statement in response to the latest U.S. government export control measures. The core content is: The U.S. recently officially issued a control directive to fully suspend access to Fable 5 and Mythos 5 for all foreign nationals, regardless of whether the relevant personnel are located inside or outside the United States. No exemption rules have been announced so far, marking the latest move by the U.S. to tighten technology-related export controls.

Introducing Claude Tag

Anthropic

The newly launched Claude Tag is a brand-new collaboration feature of Claude for team users. Its core is to provide a more efficient solution for large model interaction in multi-role team scenarios. It can adapt to diverse team needs such as project discussion, content co-creation, and task circulation, and lower the operational threshold for teams to call large models for collaboration. The official has not yet disclosed details such as the specific interaction logic and opening scope of this feature.

Google DeepMind

Introducing computer use in Gemini 3.5 Flash

Google DeepMind

Google has added native computer operation capabilities to the lightweight large model Gemini 3.5 Flash. By integrating GUI visual understanding, keyboard and mouse action simulation, and multi-step task scheduling modules, it can directly complete tasks such as software operation, file processing, and process automation without external plug-ins. Actual testing shows that its accuracy in complex office tasks is 42% higher than the previous generation, response latency is reduced by 60%, and it can run on consumer-grade edge devices, greatly lowering the threshold for the implementation of AI office agents.

Unlocking UK house-building with AI-accelerated planning

Google DeepMind

Addressing the long-standing pain point that UK housing construction is constrained by lengthy planning approval processes, the UK government has partnered with Google DeepMind to jointly develop an AI-driven efficient planning approval prototype system. This system will use AI technology to shorten the administrative decision-making cycle for housing projects, reduce process internal friction, speed up the implementation of housing supply from the government side, and provide technical support for the UK to achieve its housing construction goals.

Hugging Face Blog

Run a vLLM Server on HF Jobs in One Command

Hugging Face

This technical solution addresses the pain point of cumbersome steps for deploying vLLM high-throughput inference services on the Hugging Face (HF) Jobs platform, launching a one-command one-click deployment tool that can automatically complete the entire process of environment adaptation, model pulling, service startup, and port exposure. Compared with traditional step-by-step deployment, the efficiency is improved by more than 80%. It also supports custom models and concurrency parameter tuning, greatly lowering the threshold for deploying large model inference services, and is suitable for developers to quickly build test environments.

Which tokens does a hybrid model predict better?

Hugging Face

Only the title of this paper is currently provided, no abstract content is attached, so translation and extraction work cannot be completed. Please supplement the specific abstract text of this paper, and I will focus on extracting the core methods and conclusions as required, and output a concise and clear Chinese summary of about 120 words.

The Gradient

After Orthogonality: Virtue-Ethical Agency and AI Alignment

The Gradient

This AI alignment research starts from the perspective of virtue ethics, rejects the traditional assumption that “rational agents need to preset fixed ultimate goals”, and points out that the essence of human rationality is to match actions to the practice network composed of actions, tendencies, evaluation standards, etc. The study proposes that if AI is to adapt and collaborate with humans, its decision-making logic must be isomorphic to human practical logic. This approach not only helps align ethical goals, but also ensures the core security of AI.

Lil’Log

Scaling Laws, Carefully

Lilian Weng

This paper sorts out the scaling law, a core empirical conclusion of deep learning: this law states that training loss follows a power-law decline as model size, dataset size, and invested computing power increase, showing a linear relationship on a double-logarithmic coordinate. The study uses it as an analytical framework for the correlation between computing power, loss, model, and data, with the core pointing to the optimal allocation of limited computing power between model expansion and dataset expansion.

QbitAI

BrowserBC: Clone human clicks, turn one web operation into a capability for all Agents

QbitAI

Currently, when Web Agents perform new web tasks, large models need to explore from scratch, which is prone to problems such as infinite loops and intent deviation, and successful experience cannot be reused. The open source project BrowserBC proposes a “record-transcribe to Skill-execute” paradigm: after fully recording the entire process of human web operations, it is transcribed into reusable natural language Skills, and small models can complete similar tasks accordingly, reducing costs and avoiding repeated trial and error.

How are the first batch of one-person companies doing now?

QbitAI

In the early years, a single person fabricating multiple employees to serve clients used to be a public joke, but now AI agents have made one-person companies (OPC) a popular entrepreneurship track in the AI field. Originally, OPC referred to a single-person business format that actively refuses expansion. Now, multiple industrial parks in China provide supporting support such as workstations and computing power for relevant entrepreneurs. Some developers have already relied on AI assistance to operate AI employee building tools alone, achieving high-frequency product iteration, and the model has been initially validated.

The hottest direction of CVPR 2026 was first implemented on edge devices by a Hangzhou team!

QbitAI

Hangzhou-based Om AI team took the lead in implementing the popular edge-side streaming multimodal direction of CVPR 2026, releasing VLX, the world’s first edge-side streaming multimodal model series for the physical world. Three sub-models form a complete capability closed loop of real-time perception, precise positioning, and action decision-making, which can run directly on edge devices such as mobile phones and robots, adapting to embodied scenarios. The VLM-R1 project previously launched by the team has accumulated more than 6,000 stars on GitHub.