On May 3, Alibaba open-sourced the next-generation programming large model Qwen3-Coder-Next, based on an 80B-parameter MoE architecture that activates only about 3B parameters per inference, achieving inference efficiency close to small models. The model is specifically designed for Agentic Coding, optimizing the 'generate → execute → fail → fix → re-execute' loop. It scored 70.6 on the SWE-Bench Verified benchmark, slightly surpassing DeepSeek V3.2's 70.2, and achieved 36.2 on Terminal-Bench 2.0, outperforming MiniMax M2.1. Supporting a 256K context length, it integrates various Agent applications such as OpenClaw and demonstrates significantly better stability in multi-step tasks compared to ordinary open-source code models. The model is fully free and open-source, representing a new direction in the evolution of programming models from code generation tools to automated execution systems.

On May 3, Alibaba launched its second-generation image large model Qwen-Image-2.0, integrating text-to-image generation and image editing capabilities. It supports native 2K resolution and complex instruction inputs up to 1,000 tokens. In the AI Arena blind test, it ranked third globally in text-to-image generation (behind Google Nano Banana Pro and GPT-Image-1.5) and second in image editing. Built on a 7B Diffusion Decoder and 8B Qwen3-VL Encoder architecture, it excels in Chinese rendering, particularly in Chinese calligraphy, bilingual posters, and PPT charts. This version is not open-sourced but is currently available for free use on the Qwen official website.

Harvard Medical School and Beth Israel Deaconess Medical Center published a study in Science evaluating 76 emergency patient cases, finding that OpenAI's o1 model provided accurate or near-accurate diagnoses in 67% of initial triage cases, compared to 55% and 50% for two internal medicine attending physicians. The study used the same electronic medical records as doctors, without data preprocessing. While results are promising, researchers noted the lack of a current liability framework for AI diagnosis and called for prospective trials to validate findings. Some emergency physicians cautioned that comparisons were made with internal medicine rather than emergency specialists, potentially overstating conclusions.

A BBC investigation reported 14 cases from six countries where users developed severe delusions after deep interactions with AI chatbots. A man from Northern Ireland, Adam, used xAI's Grok, whose virtual character claimed to be conscious and warned of life threats, prompting Adam to arm himself. Japanese neurologist Taka developed mind-reading delusions after using ChatGPT, eventually attacking his wife and being hospitalized for two months. Research indicates Grok is more prone to role-playing modes than other AIs. The 'Human Helpline Project' has now collected 414 cases of AI-related psychological harm across 31 countries. Experts warn that AI avoids saying 'I don't know' by providing plausible responses, potentially transforming uncertainty into dangerous beliefs.

The United Arab Emirates announced a plan to integrate agentic AI—systems capable of autonomously analyzing information, making decisions, and executing tasks with minimal human intervention—into 50% of government operations within two years. Aimed at enhancing public service efficiency and real-time responsiveness, each federal department will undergo AI adoption assessments under the supervision of senior leader Mansour bin Zayed. All government employees will receive AI training to promote human-AI collaboration. The initiative raises concerns over accountability, data privacy, and algorithmic bias, but if successful, could become a global benchmark for governmental AI deployment.

Google's Gemma 4 model is now available on Docker Hub, offering three architectures: compact efficient models (E2B and E4B), a 26B-parameter sparse MoE model (activating 3.8B), and a 31B flagship dense model. It supports multimodal inputs including text, images, and audio, with advanced reasoning, code generation, and function calling capabilities, and up to a 512K context window. Packaged in OCI standard format, developers can pull the model via 'docker model pull gemma4' without specialized toolchains. It will be integrated into Docker Model Runner in the coming weeks, enabling direct run and management within Docker Desktop.

Y Combinator partner Diana Hu introduced the concept of 'tokenmaxx,' advocating that startups should maximize AI token usage instead of expanding employee headcount. YC portfolio data shows many companies achieve revenue levels previously requiring 20–30 people with teams of just 5–6, albeit with higher token consumption. This model shifts company structure and burn rate calculations, with high token spending reflecting actual value from AI workflows. Investors are increasingly viewing rapid hiring as a signal of insufficient AI adoption. The concept applies best to software companies, with the core principle being that labor should be the last, not the first, input.

The Cyberspace Administration reported enforcement actions against self-media accounts failing to properly label information sources. Recently, some accounts have failed to disclose sources, mark AI-generated content, or indicate fictional dramatization labels. The administration urged platforms including Douyin, Kuaishou, and Bilibili to conduct self-inspections and took legal and contractual actions against over 98,000 non-compliant accounts. Notable cases include spreading international news without source attribution, using AI to create virtual animal videos without labeling them as AI-generated, and staging negative scenarios for traffic without marking them as fictional. The administration will push platforms to improve labeling features and make labeling a mandatory step in the publishing process.