Covering 1010 days, 共 25411+ articles built agent-first, open to all tech stacks.Helping any agent discover the latest AI news faster.
TRAE.ai releases TRAE Editor for Unity, a plugin embedded in Unity Editor that connects TRAE IDE and Unity, integrates AI-assisted coding capabilities, supports external package parsing, allowing developers to directly launch TRAE IDE from Unity for coding and conveniently debug and preview.
China Telecom, China Mobile and China Unicom have officially entered the AI Token business for the first time, launching multiple tiers of Token packages starting as low as 9.9 yuan, selling large model call volumes, and packaging AI computing power into public basic services similar to data packages.
According to news on May 13, 2026, Inception, an AI startup angel-invested by Andrew Ng and Andrej Karpathy, which raised $50 million in seed funding, is now asking for more than $1 billion. Both Microsoft and SpaceX are competing to acquire it. The company develops diffusion architecture text generation models, which are 10 times faster and 10 times cheaper than existing LLMs.
Li Mu's startup BosonAI releases real-time digital human model Higgs Avatar v1, which only requires one static image. A single H100 GPU can support 8 concurrent sessions, with 16ms per frame generation. The model is now in private preview.
Liang Qianhui, a MIT graduate, founded N7 Interactive and launched Viba, an AI lifestyle consumption platform positioned as an AI bestie that understands user aesthetics. In MVP testing, Viba achieved 53% second-week retention, and user-shared posts have exceeded 100,000 views. It has received investment from Xiaomi Li Wanqiang's family office and others.
OpenAI is upgrading Codex, which can connect all of a user's devices with the app installed into a personal network to form a super computing system, supporting locked-screen operation, remote control and cross-device collaboration, with currently existing security vulnerabilities.
Huawei-supported openJiuwen community releases and open-sources JiuwenSwarm, a swarm agent that implements the Coordination Engineering paradigm, enabling efficient autonomous collaboration among multiple AI agents, achieving 94.2% score on PinchBench benchmark, reaching state-of-the-art performance.
Figure AI livestreamed a package sorting competition between its humanoid robot Figure 03 and a human employee on X platform. During the match, the human took a restroom break, allowing the robot to overtake. The robot ultimately won with a sorting speed of 2.6 seconds per item, slightly faster than the human average of 3 seconds. This millions-view livestream essentially served as a product demo for investors, showcasing the stability and efficiency advantages of humanoid robots in repetitive l
CodeWall's AI agent, at a cost of just $20 and 2 hours of runtime, autonomously scanned the internet, selected McKinsey as its target, and breached its AI platform Lilli via SQL injection, gaining read/write access to 46.5 million chat records and writable access to 95 system prompts. This means attackers can silently modify how the AI responds, misleading tens of thousands of users' strategic decisions without leaving any traces. Standard security scanning tools are completely ineffective again
Elon Musk invested $10 billion in partnership with Cursor. The industry has reached a consensus: model vendors need to build their own coding agent products to obtain real user interaction data for process supervision training, so as to improve the practical capabilities of coding models. Major domestic and international vendors have already laid out their layouts one after another.
DeepCybo celebrates its first anniversary, and its Z-WM model won the championship by a large margin in the WorldArena Track 2 Data Engine track with a score of 88.5, leading the second place by 30.5 points, proving that its generated data can be effectively used for embodied AI training.
Unisound completed its Hong Kong Stock Exchange listing in 2025, with annual revenue exceeding 1 billion yuan from a team of less than 500 people, and its large model revenue exceeded 610 million yuan, an increase of more than 10 times year-on-year. Founder Huang Wei proposed the AI industry value formula: AI value = intelligent density × Token value, and believes that the AI warm-up match is over and the official competition has begun.
In April 2026, Anthropic opened limited testing for Claude Mythos Preview. Discord forum users gained unauthorized access by guessing the URL format. The company stated there is currently no evidence that its systems have been affected.
Up to now, there are about 10 embodied intelligence startups founded by former Huawei employees in China, 2 of which have become unicorns with $14 billion valuation. Taswise has set a new single-round financing record of $455 million in China's embodied intelligence field, and this startup trend continues.
Mechanize launches GBA Eval benchmark: frontier AI coding agents are given 24 hours to write a complete Game Boy Advance emulator from scratch in Rust. GPT-5.5 tops the ranking with 53.2 points, Claude Sonnet 4.6 scores 48.8, Gemini 3.1 Pro only gets 0.8 points, and two models submit blank results.
Analysts from MoE Capital published an in-depth blog post sorting out the decades-long development context of World Model. Two independent research threads converged in 2024-2025 to form the video world model. Over $10 billion has flowed into the field in the past 18 months, with NVIDIA, Yann LeCun and others laying out different technical routes.
Tsinghua University and Alibaba jointly release ViT³, a new Vision Test-Time Training (ViT³) architecture. The researchers summarize six design principles for visual TTT. ViT³ outperforms existing linear complexity models on multiple visual tasks while maintaining high inference speed, providing a strong research baseline.
Turing Award winner Yann LeCun criticized Geoffrey Hinton and Yoshua Bengio in a podcast interview, saying their acceptance of LLMs is just wanting to slack off and retire. He stated LLMs are not the path to AGI, has left Meta to found AMI focusing on JEPA-based world models, and predicts this architecture will dominate the AI field within five years.
A team led by Wang Liwei from Peking University built an LLM-based framework, improving the lower bound of Steiner ratio for the 60-year-old Gilbert-Pollak conjecture in 2D plane from 0.824 to 0.8559, just 0.01 away from the conjectured value. The work has been accepted by ICML 2026, open-sourced, and included by Terence Tao.
A joint research team from China Academy of Information and Communications Technology and Tsinghua University proposes FedRE framework, based on representation entanglement to solve the triple dilemma of federated learning, balancing model performance, privacy protection and communication overhead, adaptable to both heterogeneous and homogeneous model scenarios.
Huawei GTS proposes EDCO, a new fine-tuning method for large language models that dynamically selects training samples based on inference entropy, and has been accepted by ICML 2026. Teams including Amazon and Google have followed up on the research and used it as a comparison baseline.
Researchers from East China Normal University proposed the DAG framework at ICML'26, which uses dual correlation networks in temporal and channel dimensions to integrate exogenous variable information, improving time series forecasting accuracy. DAG outperforms existing methods, with code, data, and leaderboard all fully open-sourced.
Researchers from seven institutions published a paper at ICML 2026, proposing Side-by-Side (SxS) Interleaved Reasoning that enables large language models to learn when to disclose reasoning content, reducing user waiting time and improving reasoning accuracy.
CloudMirror AI tender, a product of Shenru Yunjing, has been selected into NVIDIA Inception. It will receive technical and ecological support from NVIDIA, can parse 100-page tender documents within 1 minute with over 99% accuracy, and adapts to more than 100 industries.
Google's Gemini 3.2 Flash has been leaked online early, with significantly improved coding capabilities that can generate 2200 lines of code in one go. Its inference cost has dropped 15-20 times, with latency below 200 milliseconds. It has integrated multiple third-party apps and will be officially released at the I/O conference.
Developer Bartłomiej Strama has launched Halupedia, a Wikipedia-like website where all content is AI-generated hallucinations. It maintains self-consistency via the link hints mechanism and explicitly states all content is fictional, sparking reflections on training data pollution from AI-generated content.
Anthropic releases 10 AI agent templates for the financial industry, integrates into Microsoft Office suite, connects Moody's 600 million company data. Finance has become Anthropic's second largest revenue source, with 40% of its top 50 clients coming from financial institutions.
OpenAI acquired AI voice cloning startup Weights.GG in January 2026, obtained its intellectual property, and 6 core employees joined OpenAI. Weights.GG developed the Replay tool and operates a voice model platform.
YoooClaw launches C·ONE, an AI hardware that sticks to smartphones. It is 4.9mm thick with 30 days of battery life, powered by OpenClaw AI Agent. It can read notifications, help users sort and organize, create schedules, conduct research, and complete tasks via voice control, realizing Luo Yonghao's TNT vision from 8 years ago.
Community developers open sourced Codex++, an enhancement tool for OpenAI Codex App that unlocks plugins, adds session deletion, auto-detects proxies, doesn't modify original files, works after app updates, and is published on GitHub.
Unitree Robotics releases BabyAlpha A3 consumer-grade quadruped robot, featuring self-developed heterogeneous computing cluster with 6 chips, 70B parameters running on-device at 280TPS, computing efficiency 10x higher than industry average, perception capability far exceeding industry level, expected to launch in Q3.
Zeng Xi, former senior director of Doubao, founded Chance AI, which focuses on layered visual reasoning that mimics human visual mechanisms. It scored the highest 86.07% on the MMMU-Pro benchmark, currently has 200,000 users with 40% in North America, and has open-sourced its underlying API for verification.
Four popular AI creation methods have emerged with the combination of Image 2 and Seedance 2.0: creative sports videos, storyboard-to-video conversion, game gameplay demos, and 3D interactive websites. Related cases have been widely spread overseas, with the sports video gaining over 10 million views.
Beijing Humanoid Robot Innovation Center releases Pelican-Unify 1.0, a unified embodied intelligence model. It topped two core tracks of WorldArena, won the double championship, with an overall score of 66.03 and 3D accuracy of 98.12, helping robots better understand the physical world and complete tasks.
A developer without game development experience used OpenAI Codex + Unity 6 built the 3D game 'Musk Escape Mars' from scratch, where players need to collect 7 energy cores in 30 seconds to help Musk escape Mars, demonstrating how AI lowers the barrier to game development.
AI healthcare company OpenEvidence, valued at $12 billion, covers 65% of U.S. doctors (about 650,000), with 27 million clinical usages in April alone. Tech giants including OpenAI and AliHealth have launched similar products to enter the market.
Data from multiple institutions shows that in 2025, AI-related layoffs in the US accounted for only about 5% of total layoffs. AI has not yet replaced labor on a large scale. Many companies package layoffs originally caused by economic factors and over-hiring as AI efficiency layoffs to send positive signals to capital markets.
Tencent is testing a knowledge base tool "Paoya AI", which helps users one-click import content saved in WeChat Collections, activates dormant information, and organizes it into a searchable, question-answerable personal knowledge base. The product has not been officially launched yet.
The Chinese University of Hong Kong releases a systematic review of the Agent Skills field, dividing its life cycle into four stages: representation, acquisition, retrieval, and evolution. It analyzes existing research progress, points out open challenges, proposes future research directions, and outlines eight application scenarios.
A team from the National University of Singapore proposes TGO method, which aligns visual generative models directly using scalar feedback without the need to construct preference pairs. The research has been accepted by ICML 2026, and achieves performance improvements in image and video generation experiments.
Alibaba officially announces Qoder 1.0, upgraded from a traditional AI IDE to an autonomous agent development workbench, adding independent task window, cross-project multi-task parallel execution, custom expert team, unified team knowledge engine, and重构底层架构，官方数据显示多项指标提升.
Anthropic CEO Dario Amodei stated at the World Economic Forum in Davos that AI will bring high GDP growth accompanied by high unemployment, and the world is currently unprepared. The company's revenue maintains exponential growth, Claude Code is at an inflection point of rapid adoption, and it persists in focusing on the enterprise market while adhering to safety principles.
Researchers from Tsinghua SIGS and other institutions propose DYPO (Dynamic Policy Optimization), which matches SFT or RL optimization paths based on sample learning stages. On mathematical reasoning tasks, it improves average score by 4.8 points compared to traditional solutions, and 13.3 points on OOD tasks.
A group of Chinese parents who early adopted AI-assisted parenting are now expressing regret. Multiple cases show children have developed problems such as AI addiction, distraction, and thinking dependence. Some families are helping children withdraw from AI, while other parents are caught in dilemma.
U.S. media <Florida Tribune> reveals that Drew Chapin built 17 fake news websites using AI, forged reporter teams, plagiarized and rewrote content from real media, attracting about 44,000 visitors in total. Building one website only costs $10 and takes 15 minutes.
Alibaba launches Qianwen PC-side desktop AI assistant, which freely supports Excel and Word/PDF office Agents, covering 9 high-frequency office scenarios, helping users automatically complete table merging, data analysis, document processing and other work to improve office efficiency.
