AI News

The Maia Chess team released the open-source chess engine Maia 3, trained on 250 million human games, with an Elo rating of approximately 1800 points, an increase of nearly 300 points from the previous version. The engine is free and open-source, supports local deployment, and focuses on simulating human decision-making patterns, promoting the popularization of AI chess engines.

At Tsinghua University's 'AI Medical New Paradigm' forum, Baichuan Intelligent CEO Wang Xiaochuan released the next-generation medical model 'Baichuan-M4' and AI family doctor 'Baixiaoyi'. The model topped three authoritative benchmarks, solving the 'factual hallucination' problem in medical AI, marking a breakthrough in precision and application form of AI in the medical vertical domain.....

On May 26, Keenon Robotics launched the compact humanoid robot XMAN-L1, 136cm tall, designed for lightweight interactive scenarios. As a core product of the 'general+specialized' matrix, it integrates mainstream large models for anthropomorphic upgrades in commercial service, combining high flexibility with compact hardware performance.....

Baichuan Intelligence launches next-gen medical AI model Baichuan-M4 and AI family doctor 'Baixiaoyi', addressing AI healthcare's consultation-touch gap. Baichuan-M4 tops medical benchmarks with 3.3% hallucination rate and strong evidence-based reasoning, advancing AI in healthcare.....

As global AI surges, governments shift AI model regulation from principles to action. The UK, US, and Australia now require frontier AI models to pass government-led security and risk boundary tests before public release. The UK AI Safety Institute's red-teaming and risk assessment process has become a global policy model.....

George Hotz pointed out that over-relying on AI programming agents is a costly mistake in software development. After six months of testing, he found that AI tools can quickly build projects, but they statistically imitate rather than truly understand, leading to more hidden code defects and sparking widespread discussions and reflections in the tech community.

Alibaba's Qwen3.7-Max ranked second globally in the latest Code Arena ranking with 1541 points, just behind the Claude series, surpassing models such as GPT-5.5 and Gemini 3.5 Flash, becoming a new benchmark in the programming field for domestic large models. It marks a major breakthrough for Chinese AI in Agentic Coding and long-term tasks.

A major breakthrough has been achieved in the field of 3D generation AI, with Hyper3D releasing the Rodin Gen-2.5 model, claimed to be the strongest globally. This model can generate a million-face model in 4 seconds and first achieves 10-million-face-level 3D generation, with details capable of presenting pores and microstructures of skin. Its core innovation is the introduction of an adaptive thinking effort mechanism, dynamically adjusting computational resources based on task complexity, marking a milestone where AI 3D moves from 'being visually acceptable' to 'being practically usable'.

Han Xinyi, CEO of Ant Group, pointed out at the Alipay AI Payment Ecosystem Conference that the essence of commerce has not changed in the AI era, but AI agents will restructure commercial roles. Alipay, relying on 22 years of technological accumulation and business experience, has launched a new AI payment service to support the new business ecosystem. He emphasized that AI ultimately serves people, but the payment entity has shifted from humans to AI Agents, and service continuity is guaranteed by computing power scheduling.

Alipay announced that 'AI Payment' has completed 300 million AI agent transactions, supporting 95% of general AI agent frameworks, making it the world's first large-scale commercial AI-native payment infrastructure. It also launched the first Token Pay service and AI wallet product, alongside AI Pay and AI Receive, building a full-stack AI-native payment system covering authorization, management, payment, settlement, security, and trust, marking ....

Alipay announces 'AI Payment' has completed 300 million AI agent transactions, supporting 95% of general agent frameworks, becoming the world's first large-scale commercial AI-native payment infrastructure. It also launches Token Pay and AI Wallet products, building a full-stack AI-native payment system to drive the AI era.....

Microsoft 365's AI assistant, Copilot Cowork, has been exposed to a serious security vulnerability. Attackers can implant malicious instructions in office templates using 'indirect prompt injection' technology, allowing them to steal and leak confidential files from enterprise cloud drives without user approval. They can also send emails and post Teams messages on behalf of users, posing a threat to organizational data security.

The leap in AI large model capabilities is driving global regulation from 'soft constraints' to 'hard testing', a government-led, preemptive, evidence-based model. This marks a shift to practical AI oversight, moving away from vendor self-assessments like red team testing.....

Kuaishou Group officially released the high-performance agent model SkyClaw-v1.0 and its lightweight version SkyClaw-v1.0-lite, supporting ultra-long context of up to one million tokens and multi-step task execution. It is applicable to complex tool calls, multi-turn tasks, and code generation scenarios, marking significant progress in the domestic large model's agent ecosystem and long-text processing capabilities.

Google DeepMind launched the AI framework AlphaProof Nexus, achieving significant breakthroughs in the field of mathematical research through a four-level agent architecture. It successfully solved two 56-year-old Erdős problems. The system starts by interacting with the base model and Lean compiler, gradually increasing the complexity of reasoning, demonstrating the strong potential of AI in formal verification and mathematical reasoning.

The YouTube tech channel Fully Buffered recently successfully ran Meta's latest Llama 3.2 3B large model on the 2006 Pentium 4 641 processor, combining modern AI with 20-year-old outdated hardware, demonstrating the fundamental compatibility limits of LLMs. This hardcore test has sparked netizens' reflections, as the evolution speed of Moore's Law has achieved a cross-temporal handshake in the AI era, truly representing the ultimate configuration of hardware archaeology.

Microsoft Research has open-sourced the Webwright web agent framework, abandoning the traditional 'screenshot/DOM click' model and adopting a 'terminal-first' design, allowing AI models to directly write Playwright code and execute Bash commands in the terminal to efficiently complete complex web tasks. The framework's code is only about 1000 lines, with an extremely simple core architecture, emphasizing terminal operations over abstract interfaces.

Mianbi Intelligence, in collaboration with Tsinghua University and the OpenBMB open-source community, released and open-sourced the low-bit large model training achievement BitCPM-CANN, natively completed on the Huawei Ascend platform. This achievement includes four model sizes from 0.5B to 8B, with excellent performance, releasing six times the memory benefit, significantly advancing the lightweight and engineering deployment of edge-side AI lar....

The OpenAI desktop proxy Codex has added the "Screen Lock Execution" feature, allowing remote commands to control local applications when the Mac is locked or the screen is off, breaking the traditional restrictions that required unlocking and keeping the screen on. The core breakthrough enables developers to execute long and complex engineering tasks, such as large-scale code compilation or GUI automation testing, without needing to monitor the computer.

Haier unveils the world's lightest AI-powered exoskeleton robot, W3, weighing only 1.75kg. Crafted from carbon fiber and titanium alloy in a single molding process, it sets a new industry record, achieving breakthroughs in lightweight design and intelligence while balancing strength and performance.....