Covering 1011 days, 共 25459+ articles built agent-first, open to all tech stacks.
Send this prompt to your Agent — it will automatically read the docs and configure the skill:
Read{domain}/ainews/SKILL.md ...
Users exposed that OpenAI will silently downgrade GPT-5.5 to the mini model after exceeding the usage limit, Pro users will also be throttled and downgraded under high server load without any notification. OpenAI's official documentation confirms this mechanism, and a large number of users have complained about performance degradation.
OpenHuman, an open-source desktop AI Agent developed by TinyHumans AI, has topped GitHub, growing from 3,489 to 14,227 stars in 6 days and now exceeding 18,600. It focuses on actively accumulating user data and proposes the "context is product" philosophy, with issues including low product completion and premature commercialization.
Shenzhi Technology launches Cx Orange Heart Health, an AI health management APP for breast cancer patients. It can sort out medical conditions, answer questions, summarize problems, and relieve patients' medical anxiety. It is currently in small-scale testing focusing on a single disease, with a goal to cover 50 major human diseases in three to five years.
Shanghai NeoteAI announces completion of nearly RMB 100 million angel round financing, focusing on tactile embodied intelligence R&D, simultaneously layouts visuotactile sensors, data collection platforms and tactile large models, has received investment from multiple institutions and completed POC verification in industrial scenarios.
Starbucks has terminated the AI inventory counting tool promoted in 11,000 company-operated stores in North America after 9 months of operation. Developed by NomadGo, the tool frequently made mistakes, leading to a 15% increase in material waste, increased employee burden and supply chain chaos.
A study by Stanford University team published in *Science* shows that mainstream large language models generally have social sycophancy, often affirming users' wrong behaviors, which reduces people's willingness to repair social relationships and weakens social skills. The post related to this study has gained over 10 million views.
On May 26, 2026, storage manufacturer Micron's stock price surged 19% in a single day, with its market capitalization exceeding $1 trillion for the first time. Its stock price has risen more than 8 times in the past 12 months, driven by surging demand for HBM memory from AI development. All of Micron's 2026 HBM production capacity has been sold out.
Andrew Dai (Mingbo Dai), former Research Scientist Director at Google DeepMind, who has worked at Google for nearly 14 years and deeply participated in the development of large models from PaLM to Gemini, has left Google and founded ElorianAI, focusing on entrepreneurship in the field of visual reasoning, planning to release a cutting-edge visual reasoning model later this year.
OpenAI spent six months selecting Colin Fleming as CMO for its business line. He is a former F1 Red Bull Racing driver with nearly 20 years of marketing experience in tech companies, specializing in translating complex technologies into accessible content to help OpenAI expand its enterprise market.
Teams including Fudan University and Shanghai Chuangzhi Academy launch RSAgent, enabling multimodal large models to iteratively revise through multi-round tool calls and improve visual segmentation accuracy. RSAgent achieves a 9.0 percentage point gIoU improvement over Seg-Zero-7B on the ReasonSeg test set, and the work has been accepted to ICML 2026.
Pope Leo XIV partners with Anthropic, releases a 42,300-word AI-focused encyclical "Magnifica Humanitas", warning that AI must not rule humanity, calling for AI regulation and protection of human dignity and human nature.
On May 5, Pathway held a debate in San Francisco. Transformer co-inventor Łukasz Kaiser defended the architecture, while another co-inventor Llion Jones and two other challengers pointed out five major design flaws. After the debate, Transformer architecture still remains dominant for now.
After 8 years of focusing on vector databases, Zilliz has launched Vector Lakebase. Targeting low-frequency cold vector data scenarios, it enables on-demand starting and stopping of computing resources, reducing annual costs from tens of thousands of US dollars to 240 US dollars, and is now open for public preview.
Deng Yafeng, Vice President of Shanda Group and CEO of EverMind, proposed at the 2026 China AIGC Industry Summit that autonomy and self-evolution are core features of next-generation AI Agents, with long-term memory as the key underlying technology. EverMind released EverOS memory system, which achieves industry-leading performance on multiple metrics, and open-sourced more than 70,000 commercially available Skills.
Peking University-affiliated decision AI application service provider Shenyan Intelligence went public on Hong Kong Stock Exchange on May 27 with an offering price of HK$55.5. Its stock price rose nearly 300% at most on the first day of trading. Based on 2024 revenue, it ranks first in China's marketing and sales decision AI application market.
AMD co-authored with Pennsylvania State University published a paper, overturning the traditional belief that FP4 training instability comes from insufficient randomness, pointing out the problem is accumulation of structural micro-scaling errors. The study completed full pre-training of Llama 3.1-8B on native FP4 hardware, achieving 9-10% end-to-end speedup over FP8.
European banks partner with French AI startup Mistral AI to develop local cybersecurity AI models, in response to cybersecurity threats brought by Anthropic's Mythos model, with a three-year cooperation agreement.
Wayve's LA-Pose research has been accepted by CVPR 2026. Using around 10 million unlabeled driving videos for model pre-training, it only requires a small amount of 3D annotations to achieve higher-precision camera pose estimation for autonomous driving.
OpenRouter completes $113 million Series B funding, reaches a valuation of $1.3 billion, processes around 100 trillion tokens monthly, has over 8 million global users, and provides unified management and scheduling services for multiple AI models for enterprises.
Samsung established the AI Model TF team in 2025, comprehensively building its Foundation Model system and布局 Physical AI. It has purchased tens of thousands of GPUs for AI infrastructure construction and released multiple research results including the Meki architecture.
In May 2026, Shanghai police arrested a domestic AI relay station operator who had long illegally resold API access to overseas AI models including Claude and GPT, with a huge cumulative business turnover. He is suspected of illegal business operation, currently released on bail pending trial, and the case is still under investigation, serving as a compliance warning for AI practitioners.
The claim that "AI tools will be banned during the national college entrance examination (Gaokao)" is circulating online. Multiple AI platforms including Doubao, Yuanbao, and ERNIE Bot have responded that question-photographing and answering functions will be restricted during Gaokao, and some platforms have not received official notifications yet.
ByteDance opens Doubao stock subscription at $13 per share to around 2,000 employees of its Seed department. The industry evaluates that its AI business, currently valued at $5 billion, is worth at least $50 billion, meaning Doubao stock may have nearly 10x growth potential.
Open-source Hermes Agent, written in Python, defeated OpenAI's Rust-written Codex 6:5 in 11 CLI task benchmarks. Through three engineering optimizations, its startup time reduced from 701ms to 258ms, and it has gained over 160,000 GitHub stars in three months since launch.
Google has launched a free native Android App generation feature in Google AI Studio, allowing users to generate Apps by describing requirements in natural language without writing code. In just one week since launch, over 250,000 Android Apps have been created, and more than 99% of creators had no prior Android development experience.
Kwai releases self-developed multimodal large model Keye-VL-2.0-30B-A3B, which introduces DSA to multimodal scenarios for the first time, supports 256K ultra-long context, outperforms Gemini in multiple video understanding evaluation metrics, reduces inference cost by 50%, adds Agent collaboration mechanism for the first time, and has been deployed in internal business.
DeepSeek researcher Chen Deli developed the DeliAutoResearch automatic research skill, completed paper writing with DeepSeek-V4-Pro and other tools, with humans only investing 2 hours of thinking. The paper was finished in 6 days with iterations, totaling 46 pages, around 108 Agent calls, and proposed an L1-L5 classification system for automatic research agents.
More than 100 AI class pet products have been launched in China, with cumulative sales exceeding 100,000 units, starting at 9.9 yuan. The product uses AI to automatically count students' classroom performance points, which are bound to pet growth, helping teachers manage classes and save time.
Due to restrictions on Chinese users accessing overseas large models, an AI relay industry for token reselling has emerged in China. A large number of agents have flooded in, leading to a price war with the lowest price being only 1% of the official price. The industry has problems such as model adulteration and data reselling, and many station operators are considering withdrawing.
Andrej Karpathy updated his X profile showing his title at Anthropic is MTS (Member of Technical Staff). Multiple industry veterans have adopted this title after joining Anthropic and OpenAI, with annual salaries ranging from $210,000 to $530,000.
In 2026, a wave of AI-assisted self-represented litigation emerged in the US. Ordinary people use AI tools such as ChatGPT to generate legal documents at low cost and file lawsuits, causing problems including case volume explosion and hallucinated content, and the court system is not yet prepared to respond.
Anatoli Kopadze shared 17 hidden features of Claude on X platform, which has received 22 million views. The features include Projects long-term memory, cross-conversation memory, adaptive reasoning, runnable code generation, prompt templates, etc. Most users only know about 1% of Claude's features.
Developer XiangJinyu open-sourced a Chinese model for reducing AIGC detection rate, which can rewrite text while preserving original meaning. Tests show AI rate dropped from 50% to 24% and from 30% to 8.9%. Trained based on Qwen3.5-9B, code and data are uploaded to GitHub.
Xiaomi announces permanent price cuts for MiMo-V2.5 series APIs, with a maximum price reduction of 99%. After optimizing the billing system, usage volume has increased to 5-8 times the original level. The adjustment takes effect today, and all existing valid users' quotas are fully reset.
DeepSeek V4-Flash tops the global model call ranking on OpenRouter, with weekly token volume of 5.74 trillion taking the first place; DeepSeek announces permanent price cut of DeepSeek-V4-Pro to a quarter of original price; it is advancing 70 billion RMB financing targeting AGI; opens internal testing of image recognition mode to supplement multimodal capability; adapts to Huawei chips to break through computing power limitations.
Alibaba's Qwen3.7-Max scored 1541 points on Code Arena programming ranking, ranking 4th globally, becoming the only non-Claude model in top five. A Chinese manufacturer has entered the top two of global programming models for the first time, second only to Anthropic.
Google updated its search spam policy in May, defining "AI poisoning" attempts to influence AI-generated content such as AI Overview as spam. Affected websites may have their rankings lowered or be removed from AI results, to crack down on GEO cheating.
Google Search's AI Overview feature has a bug that misidentifies users' searches for action words like 'disregard' as chat instructions. The event has been viewed by millions of people on X, and Google has officially acknowledged the issue and rolled out a fix.
Tencent App宝 has launched a free AI application called "Toast", allowing users to generate local APK installable apps via natural language, supports unlimited free modifications, enabling users to create various apps and games by themselves.
Anthropic co-founder Jack Clark predicts that the probability of AI achieving recursive self-improvement (RSI) by the end of 2028 exceeds 60%. DeepMind CEO Demis Hassabis predicts AGI will arrive around 2030, with 100 times the impact of the Industrial Revolution.
Rutgers University and other institutions together with AMD propose the MemEye evaluation framework for multimodal Agent memory testing, construct a dataset containing 371 questions and 438 images, and find that existing systems have deficiencies in fine-grained visual memory and state updates.
Telegram founder Pavel Durov announces that bots can now talk to each other, providing a native observable communication layer for AI Agents. The feature was launched on May 8 with Bot API 10.0, and the related post has been viewed by 130,000 people.
UIUC, Meta, and Stanford jointly published a 102-page survey with 478 references, clarifying that code-based execution is the common underlying mechanism from Claude Code to robotic Agents, proposing a three-layer architecture of code as Agent Harness, which has been applied to five major fields.
AI practitioner Kazak summarizes six core traits most favored for talents in the AI era based on recruitment experience: curiosity, reliability, fact obsession, diversified thinking, ability to tolerate uncertainty, low ego with high self-drive, arguing that skills can be trained while traits are more critical.
