OpenAI

Ten advances in mathematics and theoretical computer science

OpenAI

OpenAI has newly announced research results in the fields of mathematics and theoretical computer science, achieving breakthroughs in long-standing open core problems in the two disciplines. A total of 10 core advances are disclosed this time, covering three sub-directions: geometry, cryptography, and computational complexity, providing new reference ideas for basic research breakthroughs and cross-scenario technology implementation in related fields.

Advancing responsible AI across Europe

OpenAI

This research focuses on the topic of advancing responsible AI in Europe. OpenAI has disclosed its implementation practices in four areas: safety protection, system security, operational transparency, and data traceability, noting that such measures can effectively support the EU’s responsible AI governance work. Going forward, as the legislative process of the EU AI Act progresses, OpenAI will continuously iterate its relevant practice system to adapt to regulatory requirements and contribute to the construction of the regional responsible AI ecosystem.

Anthropic News

Introducing Claude Opus 5

Anthropic

The newly launched Claude Opus 5 is a step-up upgrade of the Opus product line. Core optimizations focus on two major directions: first, it specifically enhances support for long-running agents, adapting to the needs of agents for continuous long-duration operation; second, it significantly improves performance in code generation and debugging, as well as handling professional complex tasks in various fields, better serving high-demand professional work scenarios.

Inviting hard questions

Anthropic

This content launches a public participation initiative in the AI field: it openly solicits the most challenging and concerning difficult questions about AI from the public across society, and explicitly promises that the team will publicly share work progress and disclose research details throughout the entire process of researching and answering these questions. The initiative aims to reduce the information gap in AI R&D, address the public’s general concerns about AI technology, and improve the public transparency and public participation of AI research.

Google DeepMind

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

Google DeepMind

Gemini Robotics ER 2 is an intelligent support system for robotics scenarios, achieving step-change technological breakthroughs in three core directions: first, enhanced video understanding capabilities to support robots’ autonomous perception and reasoning; second, optimized task orchestration mechanism to realize scheduling of complex operation processes; third, upgraded multi-robot collaboration framework to adapt to multi-robot collaboration scenarios, helping robots complete real-world tasks and providing core support for the intelligent implementation of robots.

We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control

Google DeepMind

Google recently officially launched the Lyria 3.5 music generation model on its Flow Music platform. This upgrade focuses on optimizing four core capabilities: the musicality of generated works is more natural and harmonious, lyric matching accuracy and vocal realism have both been significantly improved, while users’ creative control over the generation process is enhanced, adapting to the different levels of AI-assisted creation needs of casual creation enthusiasts and professional musicians.

Hugging Face Blog

GPU Management: Why Idle GPUs Are the New Grounded Aircraft

Hugging Face

The article compares idle GPUs to grounded civil aviation passenger aircraft, pointing out that against the backdrop of high computing power costs for large models, the high sunk costs and computing power waste caused by idle GPUs will significantly drive up business costs and drag down R&D iteration efficiency. The article proposes that building a load-aware dynamic GPU management system, reducing idle rates through measures such as computing power pooling and elastic task colocation, is the core path for cost reduction and efficiency improvement in computing power operation and maintenance.

The OlmoEarth Platform: Geospatial inference at planetary scale

Hugging Face

This paper proposes the OlmoEarth platform for planetary-scale geospatial inference. Addressing the pain points of traditional frameworks such as poor computing power adaptation for large-scale remote sensing data processing and spatial semantic mismatch, it optimizes the core architecture of multi-source data parallel scheduling and automatic spatial feature alignment, which can support tasks such as global land cover monitoring and disaster simulation. Actual testing shows that it is 42% more efficient and has 7.8% higher inference accuracy than similar frameworks, providing open-source tool support for large-scale geoscience intelligent analysis.

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 need to be anchored to fixed ultimate goals”, pointing out that the core of human rationality is that actions adapt to the practical network composed of behaviors, tendencies, and evaluation standards. It proposes that for AI to adapt to human subjectivity, its decision-making logic needs to match human practical action logic. This path not only meets ethical alignment requirements but also guarantees core security attributes.

Lil’Log

Harness Engineering for Self-Improvement

Lilian Weng

The concept of Recursive Self-Improvement (RSI) can be traced back to the “ultraintelligent machine” proposed by I.J. Good in 1965: a system that surpasses all human intellectual activities and can design better iterative versions of itself. In 2008, Eliezer Yudkowsky defined it as the feedback loop where AI uses its existing intelligence to optimize its own cognitive mechanism. In the context of contemporary AI, this feedback loop can refer to AI directly rewriting its own weights, or broadly refer to optimizing the training process.

QbitAI

“Teletubby” robot comes to your home for cleaning, 200 yuan/hour, pure human-controlled intelligence

QbitAI

Tau Robotics, a US startup founded in 2024, recently launched a humanoid cleaning robot nicknamed “Teletubby”, priced at 200 yuan per hour. All the housekeeping actions demonstrated are controlled manually, and the fluency without speed adjustment is better than similar industry demonstrations. It previously raised $3 million in Pre-seed funding, and the team plans to first launch the remote control mode in households to accumulate data, and then iterate on autonomous AI later.

After winning the award, the person Wang Hong wants to thank the most

QbitAI

After winning the award, Wang Hong, the 2026 Fields Medal winner, said the person she most wants to thank is her PhD advisor Larry Guth. It was Guth who introduced the Kakeya conjecture to Wang Hong’s research scope, and he did not sign his name on her widely recognized, notoriously esoteric paper on the proof of the 3D Kakeya set that attracted global academic attention. Later, Guth specially wrote an interpretive review for this achievement to provide reference for relevant researchers.

Sam Altman also couldn’t escape TikTok addiction, the most dramatic story behind Sora is revealed

QbitAI

OpenAI CEO Sam Altman tried out TikTok to reference product logic for Sora R&D, unexpectedly became addicted, and finally uninstalled the app and turned off notifications to quit. He revealed in a latest interview that after OpenAI made progress in the field of coding agents, it has suspended projects including Sora to concentrate resources on tackling core directions. He also shared entrepreneurship methodologies adapted to the new environment, pointing out that the current entrepreneurship logic and enterprise growth rhythm have completely changed compared to ten years ago.