OpenAI

Making it easier for websites to express AI training preferences

OpenAI

(LLM summarization failed, please manually add content later)

Community safety report and new protections for teens

OpenAI

This Community Safety Report and New Teen Protection Measures disclosure states that in response to vulnerabilities related to violent content generation risks, OpenAI is upgrading ChatGPT’s violence risk prevention and control mechanism: expanding the coverage of safety protection rules, moving risk intervention points earlier, adding a regular public commitment to disclose public safety information. The series of measures focuses on strengthening safety guarantees for teen usage scenarios, and overall improving the platform’s public safety protection capabilities.

Anthropic News

Announcements

Anthropic

This content is a summary of the official announcement released by AI company Anthropic, which clearly states that the company’s core positioning is an R&D institution focused on AI safety. Its core business revolves around safe and controllable AI systems, with R&D goals to build AI products that have three characteristics simultaneously: stable and reliable operation, interpretable decision-making logic, and behavior that can be accurately guided by humans, clarifying the safety-first orientation of the company’s technical R&D.

Investing in the public benefit

Anthropic

AI company Anthropic has launched an innovative public welfare arrangement: it has established a new entity, the Anthropic Public Benefit Fund, and pledged to transfer part of the company’s shares to this entity. Its special structure is designed so that the fund’s shareholding ratio will increase synchronously with the growth of the company’s size, ensuring that the dividends of the company’s development can feed back into the public sector to a greater extent. This is a new practice for commercial AI companies exploring a balance between commercial goals and public responsibility.

Google DeepMind

Gemini 4 Carbon: Our next frontier model

Google DeepMind

Gemini 4 Carbon launched by Google DeepMind is its next-generation cutting-edge large model, focusing on multimodal ultra-high performance. It has specifically enhanced capabilities in complex reasoning, million-token long context processing, and scientific computing, setting new SOTA records in high-difficulty benchmarks for mathematics, code, biomedicine and other fields, while significantly reducing carbon emissions from training and inference. It is the new core foundation for Google’s AGI R&D progress.

Genie 4: A new generation of world simulation

Google DeepMind

This research introduces Genie 4, a new generation of world model for world simulation scenarios. Its core breakthrough is overcoming the realism bottleneck of real-time world simulation, capable of outputting highly realistic real-time dynamic simulation effects that previous technologies could not achieve. It can adapt to high-fidelity simulation needs in fields such as robot embodied training, digital twin scenario construction, and interactive content generation, providing a better technical foundation for the implementation of world models.

Hugging Face Blog

nemotron-halloween

Hugging Face

Currently, only the paper title Nemotron-Halloween has been provided, and the specific content of the abstract is missing, so it is impossible to complete translation extraction and a summary of around 120 words. Please supplement and upload the full abstract content of this paper, and I will accurately sort out its core methods, innovations and conclusions, and output a concise and clear summary as required.

Granite Switch 4.1.1 — Update models like you update software

Hugging Face

The core of the Granite Switch 4.1.1 technical solution launched by IBM is to realize software-like updates for AI models. Through the design of decoupling model capability modules and supporting hot-swappable incremental updates, it eliminates the cumbersome process of full retraining and full deployment during model iteration, which can greatly reduce computing power and time costs. It is also compatible with existing DevOps systems, making enterprise model updates as simple and controllable as applying software patches.

Lil’Log

Harness Engineering for Self-Improvement

Lilian Weng

The concept of Recursive Self-Improvement (RSI) originated from I.J. Good’s 1965 “ultraintelligent machine” hypothesis: such systems can surpass humans in all intellectual activities, and can also design better machines to complete self-iteration. In 2008, Eliezer Yudkowsky clarified that its core is a feedback loop: AI uses its existing intelligence to optimize and generate the cognitive mechanism of its own intelligence. Current RSI in the AI field includes both models directly rewriting their own weights, and can also broadly refer to the behavior of models optimizing their own training pipelines.

QbitAI

Top of the rankings, we got it! Zhiyuan’s dual models dominate multiple rankings this round

QbitAI

At its innovation day, Zhiyuan launched Worken, a physical AI unit that measures the effective intelligent working hours of robots, along with its calculation formula. It also released two embodied large models simultaneously: GO-3 Preview for operation intelligence won first place in three embodied evaluation rankings, and WITA-OMNI for interactive intelligence topped four similar evaluation rankings. Zhiyuan also upgraded supporting robot hardware and operating systems, and its CTO Zhi Hui Jun put forward the judgment that “having 10,000 robots deployed on the job is the real victory for physical AI”.

AI writes operators from scratch, outperforms the peak of DeepSeek’s human-written operators: 9.8% faster!

QbitAI

The operator generation Agent ForgeKernel launched by Facewall Intelligence only relies on the mathematical definition of the operator and the hardware constraints of NVIDIA B200, without referring to DeepSeek’s existing implementation. After 42 iterations over 22.5 hours, the performance of the generated FlashMLA Prefill operator is 9.8% faster than the peak human-written version represented by DeepSeek. No manual coding was involved in the entire process, and it has now been announced as open-source.

“Super Mode” is officially launched! Lenovo Tianxi AI 4.5 full-device upgrade creates a “workplace super partner” across three scenarios

QbitAI

At the end of October, Lenovo Tianxi AI 4.5 completed full-device coverage across AIPC, AI tablets, and AI phones, positioned as a “workplace super partner”. Relying on cross-device collaboration and integrated software and hardware capabilities, it has upgraded three major office scenarios: the multi-app integrated Super Mode supports executing complex office tasks with one sentence and optimizes core office skills; the reconstructed AI meeting function supports real-time transcription and AI participation in brainstorming, bringing users an efficient and borderless intelligent office experience.