The first AI remote debugger AixProbe is now open source. It enables AI to directly operate embedded hardware via the MCP protocol, comes with a debugging skill package, is compatible with various architecture chips, and supports WiFi 6 remote debugging and flashing.
Alibaba released Qwen3.7-Max, the flagship Qwen large language model, at the 2026 Alibaba Cloud Summit. It ranks first among Chinese domestic models in global blind test of large models, leads in multiple evaluations, has optimized Agent capabilities, and can independently complete complex programming and other tasks.
Alibaba Cloud launches Panjiu AL128 super-node server at 2026 Alibaba Cloud Summit, featuring T-Head's first self-developed Zhenwu M890 AI chip. The Zhenwu series has shipped 560,000 units cumulatively, supporting trillion-parameter large model running on a single node, and has served over 400 customers.
Meitu leads the millions of dollars angel round financing for Chance AI, with follow-on investments from NYX Ventures and Alibaba-affiliated investment institutions. Chance AI has built the world's first Visual Agent, with about 200,000 cumulative users, ranking first on MMMU-Pro benchmark test surpassing human baseline.
26-year-old founder Odin (Zhang Haotian) founded Valhalla Tech in Hangzhou, secured tens of millions of RMB in Series A funding led by Five Elements Capital. The team is developing a full-modal AI4S foundation model, aiming to study interactions between biological molecules and support drug discovery.
DeepSeek is building a Harness team to benchmark against Anthropic's Claude Code, developing the code agent product DeepSeek Code, and opening product manager and R&D engineer positions based in Beijing.
Zhenyue Technology, founded by former Huawei team, completes $10 million series A angel round financing, with multiple institutions as investors. The company focuses on AI interactive video and will launch its global product Leadde in the second half of 2026.
AI-driven new material R&D company SynMatAI has completed tens of millions of RMB angel round of financing, exclusively invested by Ivy Capital. The funds will be used for AI R&D and industrialization of core semiconductor materials, aiming to shorten the R&D cycle of new materials.
Researchers from Yale University in the US developed the MOSAIC AI system, which divides chemical synthesis knowledge into 2498 domains to train expert models. It only requires several GPUs, achieved 71% synthesis success rate on 35 new compounds, can directly provide lab-ready detailed synthesis steps. The relevant paper is published in Nature and the system has been open-sourced.
Former Huawei executive Liu Wulong and former ByteDance executive Tao Shuai have co-founded Beta Infinity, a consumer-level embodied intelligence startup. The company has completed two rounds of financing totaling hundreds of millions of yuan, is developing its first batch of consumer products, and the pioneer experience version is planned to be unveiled this year.
On May 26, Pope Leo XIV will release an encyclical focusing on the protection of humanity in the AI era at the Vatican, inviting Anthropic co-founder Chris Olah to share the stage. The Vatican will simultaneously establish an AI committee, covering 1.4 billion believers worldwide.
Musician Hu Yanbin uses Claude Code for AI Vibe Coding to develop YanHuo, a fan-exclusive product. Many non-programmers including actors and designers are also developing products with this approach. It allows non-technical creators to independently realize product ideas and build personalized fan interactions.
An analysis in 2026 points out that Windows lags behind in multiple aspects in the AI era: poor adaptation to the Unix ecosystem, high cost of hardware AI inference, most mainstream AI tools优先支持macOS, Mac shipments grow more than twice as fast as global PCs, and Apple has launched MacBook Neo starting at $599 to enter the market.
A team from Nanyang Technological University (NTU) published the VChain framework at ACL 2026, which introduces multimodal large model reasoning capabilities to guide video generation. It does not require retraining the original video model and improves the coherence of physical laws and causal logic in generated videos, with experimental results significantly outperforming existing methods.
AI New Rank released recent popular AIGC content: The AI sci-fi short film The Patchwright exceeded 11 million views on YouTube; Keling AI 3.0's special effect sparked a global creation trend, topping App Store overall charts in 42 countries; multiple AI re-created content went viral.
Moore Threads releases MUSA ecosystem in Beijing. It has achieved full coverage of GPU software toolchain, fully compatible with CUDA operators, merged into SGLang/vLLM open-source mainline, supports automatic code migration by AI, and cloud thousand-GPU cluster has been commercially launched.
Hygon Information proposes a chip-level endogenous security solution covering four major directions including cryptography technology, has formed a complete system, entered real industrial scenarios in multiple fields, and obtained commercial cipher product certification.
A team from Griffith University and Tongji University proposed OFA-TAD at ICML 2026, advancing tabular anomaly detection from the traditional one-for-one paradigm to one-for-all. The model only needs to be trained once and can be directly transferred to unseen target datasets without retraining, leading performance in tests on 34 datasets.
In May 2026, severe AI computing power shortage occurred in Silicon Valley. Even NVIDIA's internal researchers cannot get GPUs. H100 is out of stock, H200 price rose 30% overnight, Andrej Karpathy's teaching project is also affected, a large number of SMEs and developers struggle to access sufficient computing power.
Zilliz publishes a technical article analyzing common causes of low recall in RAG systems, explaining how to locate retrieval problems through Recall metrics, introducing three optimization strategies including hybrid search, reranking, query rewriting, and methods for systematically identifying issues.
Alibaba's QoderWork launches AI design workbench Design Desk, supports generating designs via natural language, has 140 built-in style references, can directly output React + Vite front-end projects. Users say its performance is close to Claude Design with several advantages, and free quota is currently available.
Tencent opens public beta of AI design agent platform Ardot, which supports one-sentence editable UI generation, Figma file import, one-click code conversion directly to IDE, multi-person real-time on-screen review, and registered users get 1000 free Credits.
NousResearch releases Hermes Agent v0.14, which enables converting ChatGPT Pro, Claude Pro, and SuperGrok monthly subscriptions into standard API with one command, driving local development tools at zero cost. Grok 4.3 supports 100k token context window. OpenClaw releases version 5.18 fixing over 100 bugs and adding Grok support.
In 2026, there is a global supply shortage of AI GPUs including H100, with prices rising nearly 40% and delivery cycles as long as 36-52 weeks. The core causes are insufficient production capacity of HBM memory and CoWoS packaging, and large tech companies locking up capacity in advance which drives up demand.
Boston Dynamics' Atlas humanoid robot successfully lifted a refrigerator weighing over 100 lbs (45+ kg), surpassing its training load of 70 lbs. The robot learned heavy lifting through reinforcement learning combining reference trajectories, reward functions, simulation training, and real-world testing, mimicking human body mechanics when moving heavy objects.
On May 19, 2026, OpenAI co-founder Andrej Karpathy announced he has joined Anthropic's pre-training team, where he will build a new sub-team to accelerate pre-training research using Claude. This is the third core figure joining Anthropic from OpenAI in two years.
Chinese researcher Lun Wang from Google DeepMind announced his departure recently and published a blog post pointing out that existing evaluation systems cannot predict qualitative changes in model capabilities. Evaluation is the real bottleneck restricting the next leap in AI model capabilities, and a self-evolving evaluation system needs to be built.
Anthropic adds self-hosted sandboxes and MCP tunnels to Managed Agents. Sandboxes can be deployed on customer side, supporting four partner hosting providers. MCP tunnels enable secure access to local resources. Self-hosted sandboxes are now live, MCP tunnels are available for research preview applications.
Shanghai AI Laboratory develops the SU-01 model of 30B-A3B scale, which can complete 100,000-token natural language reasoning without external tools, and reaches gold medal level in difficult Olympiad evaluations including IMO, USAMO and IPhO.
Alibaba releases Qwen3.7 preview series with Max and Plus versions. Ranked #13 on LLM Arena, between GPT 5.5 and Grok 4.2, highest among Chinese models. Math #7, Programming #10, Vision #16. Solved an IMO problem in 4 minutes. Notably improved coding and visual capabilities. Available on Qwen Studio.
Xiaohongshu research team proposes HyperEyes, a parallel multimodal search agent that reconstructs underlying architecture to solve problems of low efficiency and error accumulation in traditional serial processing. In tests, it outperforms same-size open-source models by 9.9% in accuracy, and improves efficiency by 7.6 times.
Tsingshua Yangtze River Delta Research Institute releases RWAI open-source framework and the first "Real World AI Arena" for industrial AI deployment, which can shorten AI deployment cycle by over 70% and provide replicable best practices for industrial AI.
Mozilla Firefox team used Anthropic's Claude Mythos Preview to fix 423 security vulnerabilities in 1 month, exceeding the total of the previous 15 months. 271 of these vulnerabilities were discovered by Mythos, including high-risk vulnerabilities that had been hidden for 20 years.
A research team from Fudan University and ByteDance has launched GaMMA, a large multimodal model that can handle both temporal and non-temporal music understanding tasks. They also released MusicBench, the largest music evaluation benchmark to date containing 3,739 questions. Experiments show GaMMA achieves new state-of-the-art performance.
Google I/O 2026 unveiled Gemini 3.5 Flash, surpassing 3.1 Pro with 4x faster output at 289 tokens/s. The omnimodel Omni accepts any input to generate videos with physical world understanding. Antigravity 2.0 enables 93 agents to build an OS from scratch in 12 hours. Personal AI Spark runs 24/7, integrated with Google's productivity suite.
On May 19, prominent AI researcher Andrej Karpathy announced on social platform X that he has joined Anthropic, returning to LLM research and development, and still plans to continue work related to AI education in the future.