Covering 1011 days, 共 25445+ articles built agent-first, open to all tech stacks.Helping any agent discover the latest AI news faster.
At the Code with Claude developer conference in San Francisco, Anthropic gave each attendee a M5Stack Cardputer ADV mini computer. Attendees can develop various small projects in minutes using Claude Code, the device can be taken home, and the projects are open source.
MiniMax updates its desktop Agent product, launches Mavis mode, builds Team Engine multi-Agent architecture with Leader, Worker and Verifier, solves context anxiety problem, supports multi-task context isolation, and merges Token and Agent package quotas.
Mistral AI's official PyPI package mistralai version 2.4.6 has been found implanted with malicious code as part of the Mini Shai-Hulud supply chain attack. It steals credential keys from Linux systems, affecting multiple open source projects. Attackers contaminated GitHub Actions cache to obtain publishing permissions.
OpenAI's GPT-5.5 has become the first AI model to solve a problem on the ProgramBench benchmark, with its xhigh reasoning level outperforming Claude Opus 4.7, successfully blind-writing the cmatrix program from scratch, achieving a 0.5% pass rate.
Moxin Technology, founded by 00-born Chen Tianrun from Zhejiang University, has completed a new round of 100 million RMB financing. The company focuses on the world model field and has achieved delivery in multiple industrial fields, with an all-PhD team averaging under 30 years old.
AI application WinClaw launches a permanent free Token event, releasing a total of 8,000 spots from May 7 to 17. The application helps users handle office work, search and other tasks with AI, no complex configuration required.
Lin Junyang, former technical head of Alibaba's Qwen large model, has started entrepreneurship focusing on world models and embodied brains. He has recruited several core team members and is fundraising at a valuation of approximately US$2 billion, contacting institutions including Sequoia China.
Isomorphic Labs, an AI drug discovery company spun out of Google DeepMind, has completed a $2.1 billion (approximately 14.3 billion RMB) Series B financing, setting a new record for single-round financing in AI drug discovery. The funds will be used for technology R&D and advancing drug clinical trials.
9-member team Afk.Inc has raised $4M pre-seed funding, officially launched AI Agent product Bridge on May 13, 2026. It supports replicating any AI app with one sentence and can complete various tasks for users on any device.
Neuron Intelligence has been invited to join the open-source GPU-accelerated physics engine Newton as a core steering committee member, jointly defining global physical AI simulation standards with Google DeepMind, NVIDIA, Disney Research, etc. This marks the first time a Chinese company has entered the core standard definition layer in this field.
Lai Xinlu, founder of ShareAI open source community, was a guest on a podcast to analyze the three-layer framework of Agent Harness in the leaked source code of Claude Code, and shared his KB-level open source Agent toolchain. His tutorial "Learn Claude Code" has received over 50k stars on GitHub.
Cursor founder Michael Truell divided software development into three eras: Tab, Agent, and Team in his speech. Currently 75% of enterprise code is generated by AI, Agent requests increased 15 times, and 30% of PRs within Cursor are completed autonomously by Agents.
Elsevier, together with publishing groups including Hachette and Macmillan as well as individuals, sued Meta in the U.S. for using unauthorized copyrighted papers from pirated platforms to train its Llama large model. This is the first time a top academic publisher has sued an AI company, and Meta will defend the case based on the Fair Use principle.
Chinese auto finance company Yixin has self-developed and implemented an industrial-grade Voice Agent system, adopting a Model + Harness methodology adapted to auto finance scenarios. The connection-to-WeChat-add conversion rate exceeds 20%, it supports multiple languages and dialects, and will expand to general finance fields.
Navers lab under Einsia AI releases the Frontier-Eng Bench benchmark dataset, containing 47 real-world engineering tasks without standard answers to test the continuous iterative optimization capability of AI Agent, with GPT5.4 achieving the most stable overall performance.
In the trial of Musk's lawsuit against OpenAI, Sam Altman testified in court for the first time, stating that Musk had sought control of OpenAI, planned to pass the control to his children after his death, once proposed merging OpenAI into Tesla, and had long known and supported OpenAI's transition to a for-profit structure.
Alibaba Qwen team and Zhejiang University propose ECHO, an elastic speculative decoding method that dynamically allocates verification budget to improve LLM inference throughput under high concurrency. At BS=256, it improves Qwen3-235B throughput by 14.4%, and the paper is accepted as a Spotlight at ICML 2026.
Peking University team releases RealAppliance dataset and RealAppliance-Bench benchmark, containing 100 finely modeled household appliances covering 14 common categories, for systematically evaluating models' instruction-based home appliance operation planning capabilities.
Researchers propose the LangForce method to address the issue of VLA models relying on vision while ignoring language, improving out-of-distribution generalization. It outperforms existing methods on multiple benchmarks and preserves the model's core language capabilities.
Taobao integrates with Qianwen AI, allowing users to browse Taobao, select products and place orders directly within the Qianwen App. AI provides accurate shopping recommendations and services based on user conditions, completing an AI shopping closed loop.
Doubao Input Method Mac version is officially launched by ByteDance, providing free AI voice input function with high recognition accuracy, supporting mixed Chinese-English input, user habit learning, whisper and noise resistance, with input speed 3 to 5 times faster than typing.
On May 11, OpenBMB open-sourced MiniCPM-V 4.6, a 1.3B-parameter multimodal model. It outperforms same-size competitors, can be fine-tuned with just one RTX 4090, supports mixed visual token compression, features faster inference, and adapts to both edge-side and high-concurrency cloud scenarios.
On May 11, OpenClaw core team was interviewed at ClawCon event in Shanghai, stated that multiple large Chinese companies are already running their businesses directly on OpenClaw, the project now has over 80,000 PRs, and the team plans to launch long-term support versions to improve user experience.
The U.S. House of Representatives is investigating OpenAI CEO Sam Altman, accusing him of using OpenAI resources to inflate the valuation of companies he personally holds shares in, including fusion company Helion, demanding relevant documents by May 22. Six Republican state attorneys general have also asked the SEC to conduct a review before OpenAI's IPO.
Surging AI computing power demand has driven up memory chip prices. Samsung and SK Hynix have seen explosive profit growth, with SK Hynix's Q1 operating margin reaching 72%, surpassing NVIDIA. OpenAI has locked in its production capacity for years, and the memory shortage may persist until 2030.
On May 13, 2026, Google held The Android Show conference, launched the umbrella brand Gemini Intelligence and the new hardware category Googlebooks, partnered with five manufacturers to build a hardware ecosystem for Gemini implementation, following a high-end device route.
Tether, a company valued at over $500 billion, has launched the AI medical large model QVAC MedPsy. Its lightweight version outperforms Google's large model, and uses Baichuan Intelligence's Baichuan-M3-235B from China as its teacher model, which leads global competitors in multiple medical evaluations.
Kaiming He's team released ELF, a continuous diffusion language model with a novel design. Using only 105M parameters, 45B training tokens and 32 sampling steps, it reduced generative perplexity to 24 on OpenWebText, outperforming mainstream diffusion language models.
Thinking Machines Lab, founded by former OpenAI executives, released Interaction Models technology, enabling AI to listen, see and respond simultaneously. It launched TML-Interaction-Small with 276B parameters and 0.4s latency, outperforming existing models, expected to release later this year.
ByteDance open-sources DreamLite, a lightweight on-device diffusion model with only 0.39B parameters. A single network supports both text-to-image generation and text-guided image editing, generating 1024×1024 images in 3 seconds on iPhone 17 Pro, running fully locally.
FrontierX founded by a team from Zhejiang University and Alibaba launched spherical on-device robot Aura and pet-scene derivative Vex, supports custom development, plans to start crowdfunding in 2026, first-stage shipment under 10,000 units, plans to go global starting from North America this year.
Lingjiyidong, incubated by Ningbo Xiaojiang IoT, launched version 1.0 of China's first one-stop AI Agent platform for smart hardware, providing vertical models and full-link solutions to help hardware achieve proactive intelligent services, and has received tens of millions of investment from Plum Ventures.
Zhang Li, former COO of LimX Dynamics, joined domestic embodied intelligent brain startup BeingBeyond as COO in early April 2026, responsible for corporate strategy, financing, operations and business development. The company focuses on embodied intelligent model algorithms, and its latest model ranks first globally in 6 international evaluations.
Ming Chaoping from Youware team launched Bloome, an AI collaboration product that supports Agent sharing. Users can share and call local Agents, files and cloud services in channels, there are already paid Agent scenarios, and the product focuses on usability for ordinary users.
A senior team from Tencent and Adobe launches Anijam, an AI animation creation tool, has completed tens of millions of US dollars in financing. Users can automatically complete the entire creation process by inputting natural language, supports local editing at any link, currently supports generating 5-10 minute videos, adopts a layered subscription model.
Google released a report confirming for the first time that criminal hackers have used AI large language models to discover an unknown zero-day vulnerability and prepared to launch a large-scale attack, marking the first instance of AI automatic vulnerability discovery being used in actual attacks.
The trial between Elon Musk and OpenAI disclosed that former employee Ilya still holds $7 billion worth of OpenAI shares, and President Brockman holds approximately $30 billion. Over 600 employees became millionaires through equity cashing out, and employees collectively hold shares worth over $160 billion.
PingCAP's TiDB Cloud has become the database service provider for Kimi K2.6 Agent website building service. Its virtual database architecture enables second-level creation, multi-tenant isolation and extreme low cost, and can support a scale of tens of millions of sites.
South China University of Technology releases MAGIC-TTS, the world's first text-to-speech system that achieves character-level content and millisecond-level pause control, which can precisely control local rhythm within sentences to improve key information recognition, suitable for broadcasting, teaching and other scenarios.
Since the beginning of 2026, NVIDIA has invested $40 billion in AI industry chain-related companies, with an average daily investment of approximately 2 billion yuan, covering the entire industrial chain from large models to hardware, with OpenAI receiving $30 billion.
In 2026, the interactive content track rises, multiple AI companies including Loopit focus on generating playable content from user text, their valuations grow rapidly, several have reached millions of downloads, but retention rates are generally low, some are accused of buying traffic to boost data.
Nvidia invests up to $3.2 billion in 175-year-old Corning. Corning has received large orders from Meta and Nvidia this year, with optical communications becoming its largest revenue source. As of May 11, Corning's stock has risen 591.3% cumulatively, with a total market cap of $178.5 billion.
Greg Isenberg published an article discussing AI-native companies, stating that there are only about 1000 truly AI-native companies with annual revenue exceeding $5 million globally. He proposes that companies need to restructure their architecture to be machine-readable, provides a five-step construction path, and points out that the current market has large opportunity space.
Researchers from Shanghai AI Lab and other institutions show that SFT does not inherently have poor generalization capability. Its generalization is jointly constrained by three conditions: sufficiency of optimization, data quality and structure, and base model capability, breaking previous single narrative about SFT.
Ola Dimensions, a home embodied intelligence enterprise founded in March 2026, has completed hundreds of millions of RMB in financing. The round of financing will be used to accelerate the construction of home robot data closed-loop and other work, and promote the development of trainable home robots.
Creator Liu Xiaopai has uploaded the product consultant module of his AI doppelgänger "AI Xiaopai" to Tencent WorkBuddy. This AI answers product-related questions using Liu Xiaopai's way of thinking. Tencent WorkBuddy has launched a multi-AI collaborative expert team function covering multiple fields.
Bloomberg reports that AI startup Anthropic is in early negotiations to raise at least $30 billion in new funding, with a valuation exceeding $900 billion. The deal is expected to be completed as soon as the end of May, with annualized revenue reaching $30 billion, and Anthropic plans to pursue an IPO after October 2026.
