全球 AI 新闻日报 

Meituan leads Series A funding in AI startup Mindverse (Xinzhou Technology), with total financing amounting to nearly $50 million. The company builds a continual learning Agent model, using LoRA technology to achieve low-cost capability accumulation, and will soon open source the world's first 750B parameter Agent model trained with reinforcement learning post-training on GLM 5.1.

On June 1, Unitree Robotics' STAR Market IPO passed the review, taking only 73 days from acceptance to approval. It plans to raise 4.202 billion yuan for robot R&D and capacity construction, with multiple star capitals including Tencent and Alibaba holding shares.

General AI company VAST has completed a new round of Series A+ and A++ financing totaling nearly $200 million, and officially disclosed the technical roadmap for its world model Project Eden, which natively decouples underlying state reasoning from visual rendering, enabling a persistent virtual world that supports real-time multi-user interaction.

AI for Science startup Liangkun Technology completes hundreds of millions of RMB in angel and angel+ rounds of financing, led by InnoAngel Fund. Founded by Tsinghua PhD Lu Dingshun, the company integrates quantum computing, AI and high-performance computing to build a scientific computing platform.

On June 1, 2026, OpenAI CEO Sam Altman announced that OpenAI returns to the robotics track, led by Aditya Ramesh, lead developer of Sora, planning to first develop industrial robots and ultimately build personal consumer robots.

Trajectory has completed a $15 million seed round of financing with a post-money valuation of $115 million, with Jeff Dean and Feifei Li participating as investors. The company is not building general-purpose large models, but will create infrastructure that enables AI to continuously improve using real-world usage data, and has already secured multiple star clients.

Wangezhiyuan, founded by a team of post-2000 Tsinghua PhDs, has completed two rounds of tens of millions of yuan funding from Five Dimensions Capital and FreeS Fund. It launched edge inference engine cPilot and smart platform Amis, reducing cost of local large model deployment, expects to ship tens of thousands of devices and achieve over 10 million yuan in revenue this year.

Deep Principle releases MPA (Materials Property Axiom), a foundational material model that adopts LLM three-stage training and hybrid head design. It achieves SOTA on 40 real industrial material tasks, ranking first in 35 tasks, with average error reduced by over 14%.

Lu Ziheng's team developed AI material models such as MatterSim, used AI to exhaustively enumerate 200,000 materials, confirmed that diamond has the maximum thermal conductivity among bulk materials under normal temperature and pressure, has synthesized new candidate materials with high thermal conductivity, and accelerates material research and development.

ByteDance integrates Seedance 2.0 video generation model into Doubao Classroom, which can generate AI interactive video courses with images, explanations and Q&A from textbooks, providing ready-made courses and customized courses to help students preview and learn.

NVIDIA and Microsoft launch Arm architecture Windows AI PC equipped with RTX Spark chip at GTC Taibei 2026, delivering up to 1 petaflop AI computing power and 128GB unified memory, enabling local AI Agent operation and improving creative and gaming experience.

展开查看 1 篇相关报道

Nvidia will launch an AI laptop equipped with self-developed N1X chip at Computex Taipei. Developed in collaboration with MediaTek, it is based on TSMC's N3B process, featuring 20-core CPU, 6144 CUDA cores, 128GB unified memory, and uses ARM architecture.

Xiaomi publicly discloses the technical secret behind the 99% price cut of MiMo-V2.5 series API. Through full-link optimization of Hybrid SWA+MoE+multimodal architecture, KV Cache is compressed to 1/7, cache命中率 reaches 93%-95%, and it can achieve break-even operation.

Semiconductor research firm SemiAnalysis released a report stating that a large amount of economic value created by AI is not counted in GDP, forming "dark output", involving a wage pool of about $1.5 trillion. The GDP accounting of $7.2 trillion of US service industry does not reflect AI productivity improvement, and the Federal Reserve also admits that existing statistical data has flaws.

Hunter Bown (also known as Whale Brother), who had no prior programming experience, developed DeepSeek adaptation tool CodeWhale (formerly DeepSeek-TUI) with the help of AI coding. The project reached GitHub Trending with over 10,000 stars, and he has now joined American AI company Arcee AI.

York Yang, co-founder of Dyna Robotics, believes that viewing robots as the next stage of large models is the biggest misunderstanding of embodied intelligence. Robots involve complex real-world issues such as hardware manufacturing and supply chains, and cannot grow explosively like large models. The company chooses to first polish a set of hardware and software for deployment in real scenarios, focuses on complete workflows, and pays attention to whether it can be deployed long-term to generat

On June 1, 2026, four tech giants from Google DeepMind held a public conversation, revealing the R&D behind Gemini 3.5. They stated that knowledge distillation is used to compress large model capabilities into smaller models, Gemini Omni is a true world model, predicted that future models will be able to improve themselves, and Google may only have one product - the model itself in five years.

LibTV launches Director Console, a tool that supports 3D composition by drag-and-drop scene building, setting camera screenshots as reference for AI video generation, lowering the threshold for ordinary people to create AI videos. The platform also supports video parsing, and currently launched a member discount activity.

Zhang Luyu, a middle school dropout, founded the open-source AI application platform Dify in China in 2022. The project has gained 142,000 stars on GitHub, serves over 280 enterprise clients, completed a $30 million financing, and is moving the company to Silicon Valley to expand global business.

Agnes AI, an AI laboratory ranked 9th globally, has opened its three full-modal model APIs covering text, image and video to global developers for free indefinitely starting from June 1, 2026, lowering the threshold for development and usage.

OpenAI and Cerebral Valley jointly held a voice hackathon in San Francisco, announced four finalist projects: AI for answering surgical referral calls, AI tutor for children, AI multi-agent virtual meeting room, and dynamic UI voice system for mobile phones. Public voting is currently underway.

Reactor, founded by two former technical leads from Apple Vision Pro, has raised $59 million in funding led by Lightspeed. The company provides underlying infrastructure for developers building world models, with end-to-end latency controlled within 50 milliseconds, enabling low-code development of real-time interactive AI applications.

Endless Ark, an AI anti-aging drug company founded by Oxford PhDs, completes tens of millions of RMB seed round financing led by Monolith. Funds will be used for drug production verification and construction of two AI systems, verifying on pets first before advancing human anti-aging drug R&D.

A developer built an IM-based Personal Agent product half a year ago with 70% next-day retention and 30% monthly retention, acquired hundreds of paid seed users and generated cash flow, shared product observations, pointed out that commercialization of general personal agents is a core challenge, and recommended three high-quality vertical products.

Focus Media confirmed on May 23 that it has acquired a 51% controlling stake in Titanium Magnesium AI, a GEO (Generative Engine Optimization) focused AI startup, for tens of millions of yuan. Titanium Magnesium AI's services have been deployed in 7 industries, and the current GEO market size is 48 billion yuan with a growth rate of 68%.

OpenAI open-sourced Privacy Filter, a privacy filtering model with 1.5B total parameters and only 50M active parameters during inference. It can accurately mark various sensitive information, supports local operation, features 128K context window, and is released on Hugging Face under Apache 2.0 license.

China's first national-level AI + image recognition professional competition, the National Medical Insurance Image AI Recognition Competition, will be held in Guangxi from August to October. It has collected 134,000 high-quality image datasets, equipped with a 1000P+ computing power platform, set up 8 professional tracks, and supporting 10 billion yuan industrial fund and ASEAN cooperation mechanism.

Developer Cang He spent two weeks open-sourcing the free Codex practical guide CodexGuide on GitHub, covering multi-terminal usage tutorials, currently includes 13 practical cases, and supports community co-construction.

A growing number of plastic surgery patients are asking doctors to create AI-generated faces according to AI images. AI has templated aesthetics, some demands exceed surgical precision and even pose safety risks, raising patient expectations and affecting the plastic surgery industry.

Discord launched on-device AI age verification, and a primary school student tricked the system by drawing a beard. To balance privacy and user experience, the model is lightweight with loose thresholds, making it easily fooled by simple forgeries. Meta launched an AI content scanning solution to identify underage users, which has been launched in some countries.

Microsoft open-sourced SkillOpt framework this week, which treats Agent skill documents as trainable parameters, enables skills self-evolution analogous to neural network training, achieved or tied for optimal performance in all 52 evaluation combinations, and gained 3,300 stars in one week on GitHub.

Jiaming Song, creator of DDIM and Chief Scientist at Luma AI, has left Luma AI in June 2026. He joined the company in 2023 and promoted its technological transformation from 3D generation to multimodal. No new plans have been announced yet.

UEFA announces that Alibaba has become its official exclusive partner for AI, cloud computing and global e-commerce. The partnership covers 6 seasons of multiple top-tier events. Alibaba will use cloud + AI technology to upgrade sports event operations and improve viewing experience.

MiniMax releases M3 large model, which is China's first open-source model that simultaneously features cutting-edge coding capability, 1M context window and native multimodality, outperforms overseas models in multiple benchmarks, and is available for experience now.

Zilliz published an article stating RAG is not dead, as large language models still need to access external knowledge. The article introduces four generations of RAG upgrades: hybrid search, Agentic RAG, Agent Skills, and分流, and mentions Milvus can support the upgraded enterprise-grade RAG system.

Shopify's Q1 2026 financial report shows that order volume from AI search has increased 13 times year-over-year, and the new buyer acquisition rate through AI channels is nearly twice that of other channels. However, 95% of final payments are still completed manually by consumers. Currently, AI only has a high level of automation in the product discovery stage, and there are still protocol and responsibility division issues in the full process.

A team led by Yao Quanming from Tsinghua University published a paper at ICML 2026, proposing the LMNet method that constructs a trainable network with language models as nodes, communicates via dense vectors, and improves model performance at low additional cost.

AI accelerates vulnerability discovery, triggering a "bug-pocalypse". Anthropic's Mythos uncovered over 10,000 high-severity vulnerabilities in one month, 99% of which remain unpatched. Demand for cybersecurity talent surges, with OpenAI offering up to $445,000 annual salary for an AI security researcher.

A research team from Gaoling School of Artificial Intelligence, Renmin University of China, released a 40-page survey paper that systematically sorts out the definition, construction methods and other content of Rubrics for large language models, maintains a continuously updated GitHub project. Rubrics breaks down good answers into clear evaluation items, helping large language model training and evaluation.

AI Agent product 'Alice' scored 8.1 in Guanchabang's May ranking. Founder Luo Xiaoshan uses game-based settings and engineering design to achieve consistent behavior and personalized memory, helping non-technical users complete complex tasks. Users spontaneously share interaction stories and create peripherals.

Coze officially releases Coze 3.0, with full updates across three platforms including iOS, Android, Mac, Windows and web. It supports creating projects to gather different Agents for collaboration, accessing local Agents, creating cloud Agents, provides professional templates and skill store, and enables cross-device collaboration.