Global AI Daily News

Kuaishou plans to spin off Kling AI, seeking to raise $2 billion at a $20 billion valuation. Kling's annualized recurring revenue has reached $500 million, doubling since the start of the year, positioning it to become the world's highest-valued video generation AI product, compared to competitor Runway valued at approximately $5.3 billion.

OpenAI releases 12 official Codex use cases covering PR review, data analysis, PPT generation, game development and more. The core focus is demonstrating how to delegate rules, context, and acceptance criteria to AI for execution, rather than showcasing features. AGENTS.md serves as the project rules file, helping Codex understand work patterns and enable reusable automated workflows.

Uber deployed AI coding tools and within four months, 95% of engineers adopted them, with 70% of code generated by AI. With about 6,000 engineers costing roughly $6 million monthly, the company burned through its entire annual AI budget in just four months. Traditional enterprise software budgets assume predictable usage, but AI tools have no ceiling. What was once a 'promotion problem' has become a 'usage control problem,' presenting entirely new cost management challenges for businesses.

Peter releases Peekaboo v3 stable version with consecutive updates (v3.1.0-v3.1.2), giving OpenClaw desktop operation capabilities. The tool can identify screen elements, locate buttons, and execute clicks and inputs, upgrading AI from providing chat suggestions to directly controlling computers to complete tasks, solving the last-mile problem where AI couldn't reach the desktop.

According to The Wall Street Journal, in OpenAI's employee stock tender offer, about 75 employees each sold up to $30 million worth of company shares. After the transaction, OpenAI's valuation reached $500 billion, becoming the world's highest-valued unlisted startup.

Claude Mythos achieved 50% success rate on 16-hour complex long-horizon tasks in METR benchmarks, directly exceeding the upper limits of existing evaluation frameworks. AI has crossed the threshold of autonomous action, with security workloads of 3 weeks equaling a full year of team effort and attack timelines compressed to 25 minutes, suggesting the 2027 AGI singularity may arrive ahead of schedule.

Netizens discovered a data leakage vulnerability in DeepSeek: inputting specific control tokens can retrieve random training conversation records. Experts analyze this is caused by bypassing internal control tokens during supervised fine-tuning, with the model randomly continuing generation from training samples. Gemini exhibits similar phenomena. This behavior is not a bug, but a direct display of the model's post-training data.

Anthropic's valuation approaches $1 trillion, with its IPO set to potentially surpass SpaceX as the largest in history. Its ARR surged to $45 billion within five months, and co-founder Jack Clark predicts AI will achieve self-improvement by 2028, marking the start of an intelligence explosion countdown.

In May 2026, Silicon Valley VCs including a16z and Sequoia are collectively betting on the AI-native service (AINS) track. Multiple unicorns have emerged, covering vertical fields such as law, finance, and manufacturing, with innovated business models charging based on results.

Two scholars, Feng Yao (soon to join Tsinghua University) and Liu Miao (currently assistant professor at Tsinghua University), have jointly started a business to develop a new Human-Centric embodied model paradigm, enabling robots to understand human behaviors, intentions and needs, with the goal of bringing robots into ordinary people's home living scenarios as soon as possible.

AI chip company Cerebras has restarted its IPO, after increasing the offering price, its valuation reaches $35 billion and it will raise nearly $4.8 billion, making it the world's largest IPO so far in 2026. OpenAI supports it through deep cooperation including a over $20 billion procurement agreement, and Cerebras achieved net profit of $87.9 million in 2025.

Laihua launches Doratoon, an AI comic drama platform that integrates the entire creation process. Users can generate complete comic dramas with one click by inputting text, it provides two modes: quick and professional, output can be done in 3 to 6 minutes, lowering the creation threshold.

After merging xAI into SpaceX, Elon Musk has filed trademark applications for SpaceXAI, planning to integrate AI with space infrastructure, and intends to launch satellites to build orbital AI data centers with an annual new computing capacity of up to 100 gigawatts.

Multiple studies show that mainstream large models including GPT-3.5, GPT-4, Claude-3.5, etc., experience a significant drop in accuracy on analogical reasoning tests with modified problem forms, performing even worse than 7-year-old children, with their analogical reasoning ability far weaker than that of humans.

Domestic AI music model Mureka has gained recognition from enterprise users. Many enterprises have switched from Suno to Mureka. Its V8 version ranked first globally in both vocals and instrumental in international evaluation, serving over 8,000 enterprise customers worldwide, with cooperating clients achieving significant business growth.

Zhijian Dynamics, Peking University and The Chinese University of Hong Kong jointly propose LaST-R1, a new embodiment large model paradigm that integrates latent space physical reasoning into reinforcement learning loop for the first time. It achieves 99.9% average success rate on LIBERO benchmark, and outperforms current SOTA by 22.5% on real robot tasks.

China's AI video Agent track is growing rapidly. Leading platform Creati reached over 10 million global users within one year of launch, with ARR hitting $20 million. Founders worry that large tech companies will sink from model layer to application layer to seize market share, and there are divergences on time window and development path.

Century-old time-honored brand Tongrentang Beijing partners with multiple platforms under Alibaba ecosystem to launch a one-month marketing campaign, combining AI interaction and multi-scenario marketing to attract young users. During the campaign, Taobao Medicine's sales nearly doubled year-on-year, with new users accounting for over 80%.

Researchers proposed PRISM, an efficient Test-Time Scaling framework specifically designed for discrete diffusion language models (dLLMs), at ICML 2026. It improves model reasoning capability with low inference budget, experiments show it saves up to over 6x computation and achieves significant accuracy improvement on multiple tasks.

Researchers推出Claw-Eval and Claw-Eval-Live. The former verifies Agents' real capabilities by tracking execution processes, while the latter proposes a "live" benchmark that continuously updates tasks to match real-world demands. Evaluations reveal that current Agents have obvious bottlenecks in cross-system business capabilities.

Researchers from Shanghai Jiao Tong University and Qwen team found that the bottleneck of multimodal large language models' STEM visual reasoning currently lies in visual perception rather than reasoning ability. They proposed a new method called CodePercept, and the paper has been accepted by CVPR 2026, with significant performance improvement.

Yao Shunyu, a former theoretical physics researcher, moved from Anthropic to Google DeepMind in 2025, participated in the development of Claude 3.7, Claude 4.5 and Gemini 3, and believes that the era of AI individual heroism is over, and it is currently an era of collective research.

FitX AI founded by 95-born CUHK PhD Wang Jiaze completes millions of USD in funding, will build the world's first AI Native home entertainment entrance, expected to debut at CES 2027, supports空手 interaction without controllers, adapts to needs of all family members.

After testing Slock, the author concludes that current AI group chat does not work, multi-agent collaboration only generates redundant information without improving output quality. Currently it is only a geek toy that facilitates sharing local Agents, primarily providing emotional value.

NVIDIA Dynamo team discovered that the session-specific billing header in Claude Code requests breaks KV cache reuse for 52K-token context, increasing inference latency by nearly 5x. After adding a flag to strip the header, latency returns to normal.

Nvidia founder Jensen Huang received an honorary doctorate at Carnegie Mellon University's 2026 commencement ceremony. He stated that AI overturns old computing rules, brings everyone back to the same starting line, and is an opportunity for young people.

An author publishes a comprehensive beginner tutorial for Codex, the AI programming agent launched by OpenAI. The tutorial introduces its installation, features, settings, usage methods and examples. Codex has gained popularity recently due to its high cost-performance, with its ecosystem gradually improving.

An independent developer has created the AI small product TinderProfile.ai, which generates dating app avatars in 10 minutes after users upload 2-5 selfies to help users get more matches. It currently gets around 70,000 monthly visits, with an estimated annual GMV of about $300,000, serving as a case study for niche AI small products.

Security researcher Alexander Hanff discovered that Chrome silently and automatically installs the 4GB local Gemini Nano AI model on hundreds of millions of user devices that meet hardware requirements, automatically redownloads it after deletion, and did not seek user consent. Google has gradually rolled out an option to disable it.

Sam Altman calls GPT-5.5 an "autistic genius", with obvious performance advantages, consuming nearly 40% fewer tokens than Claude. Codex reached 90 million downloads in a single week, 12 times that of Claude Code. A 16-person team unsubscribed from Claude and switched to GPT-5.5, saving $32,000 per month.

Former Google DeepMind researcher Ted Xiao (now joining Bezos' Project Prometheus) reviews three eras of robot learning: existence proof, foundation model, and Scaling, sorting out the field's nearly 10-year transformation and key decisions.

guizang-ppt-skill, an open-source project by Guizang, has received over 6000 Github stars and received a major update: adds new Swiss International style, integrates GPT-Image 2.0 for adaptive image generation, supports multi-platform cover generation in different sizes, and includes 22 ready-to-use layouts.

Google DeepMind hires University of Chicago professor Alex Imas as Director of AGI Economics. He proposes that AI will lower the price of replicable labor and increase the value of relational labor such as nursing and education.

Investor Elad Gil proposes there are three tiers of technology gaps in the AI field: top AI labs lead Silicon Valley by 3-4 months, Silicon Valley leads New York by 3-6 months, and New York leads the rest of the world by 6-12 months. The AI technology accessible to the public is already 1-2 years behind the cutting edge. This view has sparked heated discussion in the tech community.

Zhejiang University, in collaboration with Tencent YouTu Laboratory, proposes the AdaMARP framework. Using a four-channel message format and scene manager, it enables immersive AI role-playing, supporting dynamic scene switching and dynamic introduction of new roles. This work has been accepted by ACL 2026.

Genesis AI releases GENE-26.5 publicly for the first time after completing $105 million seed round funding. The company launches a full-stack embodied AI system with over 200,000 hours of human operation data and human-like dexterous hand, which can perform various fine operations, sparking industry discussions.

AI-native education product VideoTutor, developed by a 20-year-old dropout and a former Google expert, has raised $11 million in seed funding. It has surpassed 50 million views on overseas social media, allows users to generate personalized interactive video explanations, and received cooperation requests from over 1,000 institutional clients on the B-end.

Researchers from East China Normal University and Fudan University proposed the Next-ToBE method at ICLR 2026. By modifying the training objective of large language models, it enables models to consider future tokens during prediction. It achieved optimal performance in 35 out of 36 experiments with lower training overhead, improving reasoning ability.

China Mobile and Volcano Engine jointly launched a confidential large model service model at the 2026 Mobile Cloud Conference, providing end-to-end encrypted AI services for government and enterprise customers with fully encrypted data that cannot be leaked. Seedance 2.0 is already online, and Doubao Large Model 2.0 will be launched soon.

In May 2026, the AI industry formally established the concept of "AI Agent Harness", which is the peripheral infrastructure wrapping large language models, consisting of 12 core components. Optimizing Harness can significantly improve the production performance of AI agents.

A research team from the University of Bristol, Shanghai Jiao Tong University and Sreal AI has released the CreativeGame framework, which solves the problems of core gameplay forgetting and score inflation when large models generate games, enabling AI to continuously iterate around game mechanics and generate truly creative new games.

Researchers from Zhejiang University and the University of Edinburgh proposed a new indicator NCB, which judges the robustness of large model knowledge by detecting the consistency of the model's related knowledge neighborhood, can predict the judgment stability of large models under interference. Experiments show that models with high NCB have a smaller accuracy drop.