OpenAI

Launching Health in ChatGPT

OpenAI

OpenAI has officially launched the dedicated health module for ChatGPT, which is currently only available to eligible users in the United States. The module allows users to securely bind their personal medical records and health data from the Apple Health platform, and leverages LLM capabilities to output health insights more tailored to individual user conditions, helping users understand their own health status more clearly and systematically.

Building AI infrastructure with the Effingham County community

OpenAI

OpenAI officially announced the launch of the “Camellia Project” in Effingham County, Georgia, to build AI infrastructure. The project has made four local implementation commitments: adopting responsible energy utilization plans, investing special community funds locally, creating local jobs, and opening access to the Codex model for local residents, to explore a mutually beneficial implementation model for AI infrastructure and local communities.

Anthropic News

Introducing Claude Opus 5

Anthropic

Claude Opus 5 is a stepwise iterative version of the Opus-tier LLM, with core targeted optimizations for long-running agent scenarios, which can stably support agents to perform complex tasks over extended periods. At the same time, this version also has significant performance improvements in code development and professional work processing scenarios across various fields, providing stronger capability support for high-difficulty development, professional office work and research needs.

Inviting hard questions

Anthropic

The core content of this AI field project document titled “Inviting Hard Questions” is: the relevant research team is openly soliciting the most concerning and challenging questions about the AI field from the whole public, and has made a public commitment that in the entire process of responding to and answering these questions in the future, it will fully disclose all process details of the relevant research work to ensure the R&D process is transparent and traceable.

Google DeepMind

Accelerating the frontiers of scientific discovery: Google’s $40M commitment to the Genesis Mission

Google DeepMind

To accelerate cutting-edge scientific discovery and expand the boundaries of research, Google announced it will invest a total of $40 million worth of AI tokens and cloud computing power credits to the Genesis Mission. These resources will be allocated to relevant research teams for calling Google’s AI capabilities to carry out interdisciplinary research, which can greatly reduce the computing power cost of AI-assisted scientific research and effectively improve the efficiency of research breakthroughs in multiple fields.

Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber

Google DeepMind

Google officially released three new lightweight large models in the Gemini series this time, namely Gemini 3.6 Flash, 3.5 Flash-Lite and 3.5 Flash Cyber. Among them, 3.6 Flash balances multimodal capabilities and inference efficiency, focusing on low-latency, cost-effective general cloud scenarios; 3.5 Flash-Lite is extremely lightweight, suitable for offline deployment on end-side devices; 3.5 Flash Cyber is a vertically optimized version for the cybersecurity domain, which can meet customized performance requirements for relevant scenarios.

Hugging Face Blog

Bringing Nunchaku 4-bit Diffusion Inference to Diffusers

Hugging Face

This work implements the adaptation and integration of the Nunchaku 4-bit diffusion inference solution with the Diffusers library. It adopts a perception-aligned quantization strategy, is compatible with mainstream diffusion architectures, and adapts to various types of software and hardware operating environments. Actual tests show that compared with native FP16 inference, VRAM usage is reduced by more than 50%, inference speed is increased by more than 30%, and the generated image quality has almost no degradation, greatly reducing the hardware threshold for diffusion model deployment.

The State of Simulation for Physical AI: An Overview

Hugging Face

This review sorts out the current development status of the physical AI simulation field: it divides existing technologies into three paradigms: first-principles-driven physical simulation, data-driven neural simulation, and hybrid simulation combining the two, compares the adaptability of different paradigms in scenarios such as rigid bodies, fluids, and flexible bodies, points out that the core bottleneck is the difficulty in balancing simulation accuracy, computational efficiency, and generalization, and clarifies the optimization direction for embodied intelligence and physical robot implementation.

The Gradient

After Orthogonality: Virtue-Ethical Agency and AI Alignment

The Gradient

This AI alignment research from the perspective of virtue ethics refutes the traditional assumption that “rational agents must have fixed ultimate goals”, pointing out that human rational actions originate from adapting to the practical network composed of actions, evaluation standards, etc., rather than pointing to specific goals. It proposes that to achieve AI-human collaboration and meet alignment requirements, the AI decision-making logic needs to match human practical action logic, taking into account both ethical alignment and security needs.

Lil’Log

Harness Engineering for Self-Improvement

Lilian Weng

Recursive Self-Improvement (RSI) was first proposed by scholar Irving John Good in 1965, with the core idea that superintelligent systems can surpass human intelligence and autonomously design better agents to complete self-iteration. In 2008, Eliezer Yudkowsky clarified its essence as a feedback loop: AI relies on its existing intelligence to optimize its own cognitive architecture. Currently, there are two implementation paths for this technology: one is for AI to directly rewrite its own weights, and the other is the broader optimization of its own training pipeline.

QbitAI

100% Open Source! Andrew Ng built a personal desktop Agent

QbitAI

Andrew Ng’s team has launched the open source desktop AI Agent OpenWorker, which uses the MIT license, currently supports running on Mac, and the Windows version will be launched soon. The product focuses on openness, local-first, privacy protection, and model agnosticism, and can be connected to various mainstream closed-source LLM APIs and local open source LLMs, with data stored locally by default. Unlike chat-based AI, it can automatically execute instructions across files, calendars, and office software to deliver actual work results.

US embodied intelligence is not mature either! PI: Why do Chinese companies always have to be the “Chinese version of XX” | RSS 2026

QbitAI

This article is an observation from the RSS 2026 conference: First, leading North American embodied intelligence enterprises are far from mature, and there are no foundational achievements in the field yet. China has an advantage in the robot hardware end, so there is no need to benchmark against overseas companies and call themselves the “Chinese version of XX”; Second, world models have been widely hyped in China, but top conferences still focus on VLA discussions, the core reason is that the top conference review cycle is long, and the achievements presented at the conference were mostly launched more than half a year ago, lagging behind the rapid iteration of the industry.

Beat Fable 5 at half the price? Opus 5 test results are explosive, netizen: I almost fell off my chair

QbitAI

Anthropic’s newly released LLM Opus 5 has tied or even outperformed Fable 5 in multiple hardcore benchmarks, priced at only half of the latter, and its performance is more than twice that of the previous generation Opus 4.8. Its code and 3D generation capabilities are outstanding, even if 80% of the supporting system prompts are deleted, the effect does not decrease. Actual tests show that its generated games and animation effects far exceed expectations, and netizens evaluate its performance as comparable to Fable 6.