Global AI Daily News

Flowith team released Matrix, a product focused on multi-Agent collaboration to operate an 'unmanned company'. After users set a goal, a CEO-level Agent coordinates research, production, and analysis departments to automatically complete tasks. Matrix achieved 95.45% accuracy on GDPval-Bench, setting a new SOTA and surpassing Claude Opus 4.7. It supports integration with various external tools and payment accounts, and is currently available for macOS.

Meituan releases LongCat-2.0, a trillion-parameter MoE model trained and inferred entirely on over 50,000 domestic computing chips, making it China's first fully domestic-compute trillion-parameter model. Its Agent and coding capabilities approach the global top tier, validating the feasibility of training world-class large models on domestic computing infrastructure.

VAST has completed a strategic A3 funding round exceeding 1 billion RMB, with investors including Geely Capital, 4399, Tanwan Games, Giant Network, and other industry players alongside multiple investment institutions. The company's Tripo product, positioned as an AI 3D foundation model, can generate engine-ready mesh models in 2 seconds, transitioning from effect generation to production tooling. VAST also launched Project Eden, a world model initiative aimed at building persistent, interactive

OpenAI has proposed granting the White House a 5% equity stake in the company (approximately $42.6 billion at its $852 billion valuation), aiming to shape regulations and mitigate regulatory risks through government shareholding and dividends to citizens, though the plan has drawn criticism from both left and right.

Four companies co-hosted AI Demo Night in Silicon Valley, 10 teams did live product demos covering multiple directions including database optimization, medical data query, game development, etc. Multiple products haven't registered domain names yet, all solve real problems from actual pain points.

Feng Weihe, former CEO of Meituan Homestay, co-founded AI company Sowii with Professor Zeyu Zheng from UC Berkeley, secured investments from Sequoia and Hillhouse Capital. It launched the AI digital life product 'Tutu' with a physical plush pendant, the first batch of 30,000 units sold out in two days, pre-orders exceeded 200,000 units, and 7-day user retention reached 73%.

JD Cloud and Coolpad jointly launched the customized AI cloud phone Fengshang 50Lite, starting from 799 yuan, exclusively available at Coolpad's JD self-operated flagship store. It is built-in with a full set of JD Cloud AI services, comes with one-year membership, supports cloud installation-free apps and cloud game hanging.

Jelani Nelson, chair of UC Berkeley's EECS department, has stepped down from academia and joined Anthropic's pre-training team, focusing on research into Claude's core capabilities. His algorithm公开课 on YouTube has exceeded 21 million views, and he previously worked as a part-time research scientist at Google.

On June 24, 2026, ByteDance's Doubao Pro officially launched with three pricing tiers, the highest at 500 yuan per month. The article points out its payment logic is comparable to Melatonin, focusing on symbolic value and social attributes, becoming a carrier for young people to express filial piety, with characteristics of Veblen goods.

Wired exposes that Meta runs a secret project codenamed "Cannes", hiring outsourced workers to impersonate minors and send over 45,000 malicious high-risk prompts to competitors including ChatGPT, inducing AI to generate non-compliant content. Meta claims this is routine safety testing, while the tested companies did not authorize this test.

Alibaba has integrated its AI business to establish the ATH business group. Its Qwen 3.7 Max large model scores Opus 4.7 in benchmark tests, ranking second among Chinese large models, with only a slight gap behind GLM 5.2. The next generation is expected to become the top Chinese large model.

Mininglamp Technology launches the open-source platform Octo, designed for enterprise organizational scenarios, which can aggregate dispersed Agents, supports six collaboration modes, precipitates organizational preference data, realizes human-AI collaborative work, and the code has been open-sourced on GitHub.

Base44 launches self-developed large language model Base1, trained on tens of millions of user interaction data, reducing reliance on external APIs like Anthropic and cutting inference costs. Its annual recurring revenue now exceeds $150 million.

Ant Group, together with HKUST and Tsinghua University, released open-source AReaL 2.0, an Agentic RL training base that enables deployed Agents to convert real interaction data into self-evolution training data without massive code modifications, accelerating Agent self-evolution.

A developer discovered a logging bug in OpenAI's programming tool Codex. Due to incorrect default logging level configuration, it writes approximately 640TB of data per year, which can deplete the lifespan of a 1TB solid-state drive. OpenAI has fixed about 85% of the issue.

Huawei's openJiuwen community open-sources AutoGenetic Memory, an autonomously growing memory engine for intelligent agents that enables persistent memory across sessions. In LoCoMo evaluation, it improves accuracy by 15% and reduces token consumption by over 60%, supporting cross-platform sharing and group memory precipitation.

Ma Xiao from Zhejiang Sci-Tech University and Li Wujun's research group from Nanjing University jointly proposed EMCES, which introduces episodic memory into controllable diffusion models to guide high-quality sample synthesis for reinforcement learning, reducing storage overhead by about 8000 times and time overhead by 25.5 times. The paper has been accepted by ICML26.

Anthropic officially announces global launch of Claude Fable 5, discloses security test results showing 8 models including China's Kimi K2.7 have the same vulnerability, indicating Kimi reaches equivalent technical level. Multiple North American tech companies have adopted Chinese large models to cut costs.

Show 1 related articles

Anthropic officially announces the return of Fable 5, available from July 1 to July 7 with 50% weekly usage limit for all subscription plans. Most coding tasks are supported after updated safety measures, pricing is about twice that of Opus 4.8.

FaceMind launches Ego-NeuroLoop data paradigm, which synchronously collects four types of signals related to human movements. It is equipped with NeuroMatrix hardware to reduce acquisition thresholds, and NeuroBooster processes raw data for training better world models for embodied AI.

Meta is building a cloud business called "Meta Compute" to sell AI computing power and model access. After the news broke, Meta's stock rose more than 10%, market cap increased by nearly $98 billion, while competitors CoreWeave and Nebius saw sharp stock price drops.

Andrew Ng posted on X proposing a new Loop Engineering framework, dividing software development into three layers of loops: AI autonomously iterates code, developers correct product direction, and real market feedback adjusts judgments. The three loops become slower and more important in sequence.

Verily completes $300M financing and spins off from Google's Alphabet. Alphabet remains a minority shareholder, and Verily will focus on its AI health assistant product Verily Me, entering commercial validation targeting a health data business closed loop.

OpenAI's Codex forced two quota resets for users within 48 hours from June 29 to 30, caused by multiple overlapping issues leading to abnormal token quota consumption. OpenAI has now fixed the issues and provided compensatory resets for users.

Philo AI, an AI startup led by Huawei "genius boy" Zhang Jiasheng with all core members holding PhDs from Tsinghua and Peking University, has completed a nearly $10 million Series A funding. The capital will be used for R&D of World Life Model, creating real-time interactive digital life, targeting to achieve video-modality AGI.

Within a week, four core AI talents (including Nobel laureate John Jumper) left Google for OpenAI and Anthropic, causing Alphabet to lose hundreds of billions of dollars in market value. Google launched AI agent Gemini Spark, which can automatically complete tasks across Google apps on the cloud, only available to Google AI Ultra users paying $100/month.

NVIDIA open-sources ASPIRE, an embodied intelligent robot skill library that enables robots to accumulate reusable experience from trial and error, boosting success rate from 20% to 92% on Robosuite tasks, proposing a new continuous learning paradigm for robots.

Four months after OpenAI shut down Sora, Google released Nano Banana 2 Lite image model and Gemini Omni Flash video generation model, covering different user needs to form an end-to-end AI creation family bucket from image to video, with Gemini Omni Flash currently in public preview.

Rokid launches the world's first AI-native AI smart glasses operating system AIOS at Rokid Open Day 2026, releases matching interaction standard AIUI, and launches the industry's first AI agent app store, with over 440 agents currently available.

OceanBase officially releases Lakebase, its lake-house integrated AI database. It unifies data formats, computing paths and governance boundaries, supports multi-mode tables, hybrid search and open computing, provides elastic and scalable data infrastructure for AI Agents, with multiple performance metrics outperforming industry competitors.

Researchers from multiple institutions including University of Science and Technology of China published a review, sorting through more than 500 papers, categorizing generation consistency issues of diffusion models into three types, analyzing existing optimization methods and evaluation difficulties, and pointing out core challenges such as constraint conflict that need to be solved in the future.

AFAC2026 Financial Intelligence Innovation Competition has kicked off, co-organized by Ant Group and other institutions. It sets four competition questions from real financial scenarios, exploring industrial implementation of large models under real constraints, and has attracted nearly 50,000 participants so far.

A tester had LongCat 2.0 independently plan and develop a Tetris H5 game with five gameplay modes in Codex. The first version was released in about 50 minutes, and full functionality was completed after multiple rounds of polishing. Token consumption was around 4M, demonstrating complete project development capability.

Fudan University NLP Laboratory proposes SciAgentGym, which includes the SciAgentBench benchmark and SciForge training method, designed to evaluate and improve large language models' execution capability in multi-step scientific workflows. Evaluation shows current large models experience significant performance degradation on long-horizon scientific tasks.

95-year-old PhD founder Hongyuan "Adam" Lu founded FaceMind, focusing on general world model R&D, launched the Looped World Model architecture that can reduce multi-step prediction errors and supports hundreds of steps of long-range simulation. The company is currently raising tens of millions of US dollars with a valuation of hundreds of millions of US dollars, and has a team of around 20 people.

Zilliz publishes a technical tutorial introducing how to use Milvus' Partition Key physical isolation capability combined with AutoRAG automatic evaluation framework to solve data cross-tenant leakage problem in multi-tenant RAG and Agent systems, single Collection supports up to 4096 physical partitions.

OpenAI Chief Research Officer Mark Chen stated that scaling laws are still valid, pre-training and other methods remain the main path to AGI, believes AGI is coming soon, models will soon be able to conduct research independently, there are still challenges such as evaluation currently, but he is confident they can be solved.

Kuaishou's KeLing AI will complete $3 billion (approximately 20.38 billion RMB) financing, with a post-investment valuation of $18 billion. Tencent participates in the funding, and it is expected to launch a Hong Kong IPO within 12 months, with proceeds to be used for computing power and talent development.