Global AI Daily News 

DeepSeek researcher Deli Chen publishes his second paper co-authored with AI, focusing on continual learning and self-iteration. After iteration, his autonomous research framework DeliAutoResearch SKILL improves its simulated peer review score from 6 to 8, with enhanced autonomy.

In May 2026, a study by UC Berkeley involving over 95,000 undergraduate students and a study by Georgetown University analyzing over 370,000 college application essays found that AI use causes problems of capability hollowing out and ideological homogenization, pointing out that the real dividing line in the AI era is establishing one's own baseline of judgment before using AI.

Weijie Su, a member of the 'second golden generation' of Peking University School of Mathematical Sciences and the first Chinese winner of the COPSS Presidents' Award (top statistics honor) in 14 years, has officially announced joining OpenAI. He has also been promoted to full professor at Wharton School, and will conduct AI model training research at OpenAI.

Wharton Business School professor Ethan Mollick used Claude Opus 4.8 to independently develop the "The Veil of History" reincarnation simulator website. It calculated that approximately 117 billion humans have been born throughout history, with 81% born before 1650, and supports random life simulation. Claude Opus 4.8 currently tops multiple AI benchmark tests.

Fei-Fei Li's team, in collaboration with Stanford University and other institutions, has released the GPIC dataset, containing 100 million image-text pairs with a total of 28 trillion pixels, addressing issues such as outdated benchmarks. It is now freely available, accompanied by a new evaluation protocol for research.

An evaluator conducted a head-to-head comparison of four popular large models including Claude Opus 4.8, Gemini 3.5 Flash, GPT-5.5, and Qwen3.7-Max across seven real-world tasks, finding Claude Opus 4.8 is the strongest overall, and recommends choosing models based on specific use cases.

The Serpent in the Grove, the Caribbean winner of the 2026 Commonwealth Short Story Prize, was confirmed 100% AI-generated by tools including Claude. It was selected from 7806 submissions, beating 7805 human writers.

NVIDIA, together with Tsinghua University and other institutions, released Gamma-World (γ-World), which solves the expansion problem of existing multi-agent world models, reduces computational complexity from quadratic to linear growth, supports zero-shot generalization to four-player scenarios, and topped the HuggingFace daily ranking.

Shanghai AI unicorn MiniMax submitted A-share listing tutoring备案on May 29, kicking off its A-share IPO process with CITIC Securities as advisor. Its stock surged 409% from issue price to HK$840, with market cap reaching HK$263.4 billion. ARR grew over 100% in two months, serving over 1 million global clients and ~300 million users. M3 model is coming soon.

Singularity Perturbation releases AI digital life "Reyi", built on a proprietary hyper-personalized model trained with hundreds of thousands of words of personality data. In beta testing with 1,000 users, Reyi achieved 69% D2 retention and 58% D3 retention. Users treat her as a real presence rather than a tool, with 99% expressing clear willingness to pay.

TokenLX, launched by ZWinsoft, is an AI compute dispatch platform integrating 190+ models from 11 suppliers across 7 modalities, featuring intelligent routing for optimal model matching. Through its six-layer context compression technology, per-token consumption can be reduced by approximately 80%. Integration requires just two lines of code, with beta users enjoying 50% off on Qwen3.7 Max.

XATRIX launches an AI smart ring for couples, combining jewelry design with smart hardware to enable mutual health data monitoring. Pricing at $229 per ring or $419 for couples, featuring heart rate, SpO2 sensors, with no subscription required. The brand bonds health monitoring with social needs through the ring, targeting the niche couple market.

Tencent Games launches AI game creation platform "Codename Craft", enabling users to generate playable games from scratch via natural language, supporting both 2D and 3D. The platform features complete AIGC tools with over 20,000 free preset assets, lowering the barrier to game creation and allowing ordinary players to participate in game development.

IBM and Artificial Analysis launch ITBench-AA, a new benchmark testing large models in real enterprise IT environments. Alibaba's Qwen3.7-Max ranks third with 42.5% accuracy, solving problems in an average of only 37.6 interactions. Even the top-performing model currently scores below 50%.

According to a report from foreign media Axios, a company spent 500 million US dollars (equivalent to 3.3 billion RMB) on Claude in one month after giving employees unlimited access without usage caps, sparking widespread discussion online.

Fudan University and Alibaba Tongyi Wanxiang jointly propose DiffusionOPD, a new on-policy distillation method for diffusion models, enabling a single student model to simultaneously acquire capabilities in composition, text generation and aesthetic optimization, outperforming existing methods on multiple tasks.

On May 29, 2026, DeepSeek began limiting the number of re-generations and modifications in regular conversations and expert mode. This is a temporary measure to cope with user growth and computing power pressure. The restrictions are expected to be lifted after new Huawei Ascend cards are deployed and capacity is expanded in the second half of the year.

Cheng Mengqi's SophiaPro completed multiple product pivots in 300 days: from vertical Agent to B2B to C-end, launching TryClico browser extension and Invoko.ai desktop client. She summarized 10 key lessons from AI app entrepreneurship, core takeaway being the structural mismatch between Agent product model and B2B business model, and that software has no shortcuts—it consists of countless details. Key insight: build for yourself first, then expand outward.

SurfaceAI and OpenBMB open-source Ultra-FineWeb-L3 (600B tokens) and UltraData-SFT-2605 (10M+ SFT samples), both L3 synthetic datasets. This is China's largest open-source Chinese pretraining dataset and the first large-scale SFT dataset covering both deep thinking and non-thinking modes. 1B models achieve 2B-level performance through high-quality data, making data quality the true moat in on-device AI scenarios.

Omdia forecasts global AI infrastructure capex to exceed $600 billion in 2026. SenseCore has been identified as a pioneer of the "Intelligence Refiner" paradigm through its five-layer architecture, achieving efficient conversion from computing power to productive AI capabilities.

FaceMind pioneered low-frequency token degradation research, publishing at EMNLP 2025 and ACL 2026 Oral. Anthropic later validated the approach but only at word level—sentence-level methodology remains untapped in industry.

China's three major telecom operators launched Token packages, marking a strategic shift from "connectivity providers" to "AI computing portals." China Telecom aims to "remodel its business through Token operations," China Mobile prioritizes infrastructure, and China Unicom focuses on Agent and AI cloud services. Operators are transitioning from building to selling computing power, though challenges remain in user comprehension, benefit transparency, and high-frequency usage scenarios.

Doubao deepens integration with Douyin Life Services, enabling in-app group buying, payment, order tracking, and QR code redemption covering dining, movie tickets, and accommodations, eliminating the need to switch to Douyin for a seamless consumption experience.

Meta has offered top AI researchers compensation packages approaching $100 million. Epoch AI research shows elite AI researchers' annual salaries exceed $10 million—200 times that of postdoctoral researchers. The underlying driver is the "superstar effect": model improvements multiplied by 900 million users create exponential amplification. With the AGI arms race pot potentially worth trillions, $100 million is merely a reasonable insurance premium.

Shishi Technology launches a domestic TOKEN optimization factory, integrating NVIDIA GPUs and domestic AI chips, achieving 30%-50% throughput improvement and 40% token cost reduction through intelligent scheduling and inference optimization, with 99.9% high availability, advancing large-scale development of China's AI computing ecosystem.

Baidu has shut down AI video creation platform "Huixiang" (in April) and will close "Baidu Miaobi" (on June 15), with services integrated into AI creation platform "Dujia". Huixiang was built on Baidu's self-developed MuseSteamer video generation model, which topped the VBench global ranking. This consolidation helps reduce operational costs, unify branding, and accelerate commercialization.

Zuma Innovation completes angel+ financing round, led by SenseTime Guoxiang Capital with participation from CDH Investments and Freak Ventures. Founded roughly 6 months ago, the company's valuation doubled in about 30 days. The startup is developing a consumer-grade spatial camera using 3DGS technology, enabling users to capture 3D worlds as easily as taking photos, drastically reducing costs from tens of thousands of yuan to an affordable level for ordinary consumers.

ByteDance is attracting PM talent in the AI era. Doubao reached 100M DAU in under 2.5 years, with 227M MAU in Q4 2025, while Volcano Engine holds 49.5% MaaS market share. With its 'Always Day 1' culture, three-carriage architecture, and bold talent approach, ByteDance enables rapid iteration, making it an essential choice for PMs seeking to build outstanding AI products.